OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

POEMS THAT SOLVE PUZZLES

OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

POEMS THAT
SOLVE PUZZLES
The History and Science of Algorithms

Chris Bleakley

1
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

3
Great Clarendon Street, Oxford, OX2 6DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© Chris Bleakley 2020
The moral rights of the author have been asserted
First Edition published in 2020
Impression: 1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2020933199
ISBN 978–0–19–885373–2
Printed and bound by
CPI Group (UK) Ltd, Croydon, CR0 4YY
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

For Eileen - thank you

OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

algorithm, noun:
A process or set of rules to be followed in calculations or other
problem-solving operations, especially by a computer.
The Arabic source, al-Kwārizmī ‘the man of Kwārizm’ (now
Khiva), was a name given to the ninth-century mathemati-
cian Abū Ja’far Muhammad ibn Mūsa, author of widely trans-
lated works on algebra and arithmetic.
Oxford Dictionary of English, 2010
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Foreword

This book is for people that know algorithms are important, but have
no idea what they are.
The inspiration for the book came to me while working as Director
of Outreach for UCD’s School of Computer Science. Over the course
of hundreds of discussions with parents and secondary school students,
I realized that most people are aware of algorithms, thanks to extensive
media coverage of Google, Facebook, and Cambridge Analytica. How-
ever, few know what algorithms are, how they work, or where they
came from. This book answers those questions.
The book is written for the general reader. No previous knowledge
of algorithms or computers is needed. However, even those with a
degree in computing will, I think, find the stories herein surprising,
entertaining, and enlightening. Readers with a firm grasp of what an
algorithm is might like to skip the introduction. My aim is that readers
enjoy the book and learn something new along the way.
My apologies to the great many people who were involved in the
events described herein, but who are not mentioned by name. Almost
every innovation is the product of a team working together, building
on the discoveries of their predecessors. To make the book readable
as a story, I tend to focus on a small number of key individuals. For
more detail, I refer the interested reader to the papers cited in the
bibliography.
In places, I favour a good story over mind-numbing completeness.
If your favourite algorithm is missing, let me know and I might slip it
into a future edition. When describing what an algorithm does, I use the
present tense, even for old algorithms. I use plural pronouns in place of
gender-specific singular pronouns. All dollar amounts are US dollars.
Many thanks to those that generously gave permission to use of their
photographs and quotations. Many thanks also to my assistants in this
endeavour: my first editor, Eoin Bleakley; my mentor, Michael Sheri-
dan (author); my wonderful agent, Isabel Atherton; my bibliography
wrangler, Conor Bleakley; my ever-patient assistant editor, Katherine
Ward; everyone at Oxford University Press; my reviewers, Guénolé
Silvestre and Pádraig Cunningham; and, last, but certainly not least, my
parents and my wife. Without their help, this book would not have been
possible.
Read on and enjoy!
Chris
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

About the Author

Chris Bleakley has thirty-five years of experience in algorithm design.

He has taught and written on the subject for the last sixteen of those
years.
As a school kid, Chris taught himself how to program on a home
computer. Within two years, he was selling his own computer pro-
grams by mail order to customers throughout the UK.
Chris graduated with a BSc (Hons) degree in Computer Science from
Queen’s University, Belfast, and a PhD degree in Electronic Engineer-
ing from Dublin City University. After college, he was employed as
a software consultant by Accenture and, later, as a senior researcher
by Broadcom Éireann Research. Thereafter, he was appointed Vice
President of Engineering at Massana, a leading-edge start-up company
developing integrated circuits and software for data communications.
Today, Chris is an Associate Professor and Head of the School of Com-
puter Science at University College Dublin (UCD), Ireland. He leads
a research group focused on inventing novel algorithms for analysing
real-world sensor data. His work has been published in leading peer-
reviewed journals and presented at major international conferences.
Chris lives in Dublin with his wife and two children.
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Contents

Introduction 1
1 Ancient Algorithms 9
2 Ever-Expanding Circles 25
3 Computer Dreams 39
4 Weather Forecasts 55
5 Artificial Intelligence Emerges 75
6 Needles in Haystacks 93
7 The Internet 117
8 Googling the Web 143
9 Facebook and Friends 159
10 America’s Favourite Quiz Show 171
11 Mimicking the Brain 179
12 Superhuman Intelligence 203
13 Next Steps 215

Appendix 229
Notes 233
Permissions 241
Bibliography 243
Index 259
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Introduction
‘One for you. One for me. One for you. One for me.’ You are in the
school yard. The sun is shining. You are sharing a packet of sweets with
your best friend. ‘One for you. One for me.’ What you didn’t realize back
then was that sharing your sweets in this way was an enactment of an
algorithm.
An algorithm is a series of steps that can be performed to solve an
information problem. On that sunny day, you used an algorithm to
share your sweets fairly. The input to the algorithm was the number of
sweets in the packet. The output was the number of sweets that you and
your friend each received. If the total number of sweets in the packet
happened to be even, then both of you received the same number of
sweets. If the total was odd, your friend ended up with one sweet more
than you.
An algorithm is like a recipe. It is a list of simple steps that, if followed,
transforms a set of inputs into a desired output. The difference is that
an algorithm processes information, whereas a recipe prepares food.
Typically, an algorithm operates on physical quantities that represent
information.
Often, there are alternative algorithms for solving a given problem.
You could have shared your sweets by counting them, dividing the
total by two in your head, and handing over the correct number of
sweets. The outcome would have been the same, but the algorithm—
the means of obtaining the output—would have been different.
An algorithm is written down as a list of instructions. Mostly, these
instructions are carried out in sequence, one after another. Occasion-
ally, the next instruction to be performed is not the next sequential
step but an instruction elsewhere in the list. For example, a step may
require the person performing the algorithm to go back to an earlier
step and carry on from there. Skipping backwards like this allows
repetition of groups of steps—a powerful feature in many algorithms.
The steps, ‘One for you. One for me.’ were repeated in the sweet sharing
algorithm. The act of repeating steps is known as iteration.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

2 Introduction

If the number of sweets in the packet was even, the following iterative
algorithm would have sufficed:

Repeat the following steps:

Give one sweet to your friend.
Give one sweet to yourself.
Stop repeating when the packet is empty.

In the exposition of an algorithm such as this, steps are usually written

down line-by-line for clarity. Indentation normally groups inter-related
steps.
If the number of sweets in the packet could be even or odd, the
algorithm becomes a little more complicated. A decision-making step
must be included. Most algorithms contain decision-making steps.
A decision-making step requires the operator performing the algorithm
to choose between two possible courses of action. Which action is
carried out depends on a condition. A condition is a statement that is
either true or false. The most common decision-making construct—
‘if-then-else’—combines a condition and two possible actions. ‘If’ the
condition is true, ‘then’ the immediately following action (or actions)
is performed. ‘If’ the condition is false, the step (or steps) after the ‘else’
are performed.
To allow for an odd number of sweets, the following decision-making
steps must be incorporated in the algorithm:

If this is the first sweet or you just received a sweet,

then give this sweet to your friend,
else give this sweet to yourself.

The condition here is compound, meaning that it consists of two (or

more) simple conditions. The simple conditions are ‘this is the first
sweet’ together with ‘you just received a sweet’. The two simple con-
ditions are conjoined by an ‘or’ operation. The compound condition
is true if either one of the simple conditions is true. In the case that
the compound condition is true, the step ‘give this sweet to your
friend’ is carried out. Otherwise, the step ‘give this sweet to yourself’
is performed.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Trainee Librarian 3

The complete algorithm is then:

Take a packet of sweets as input.

Repeat the following steps:
Take a sweet out of the packet.
If this is the first sweet or you just received a sweet,
then give this sweet to your friend,
else give this sweet to yourself.
Stop repeating when the packet is empty.
Put the empty packet in the bin.
The sweets are now shared fairly.

Like all good algorithms, this one is neat and achieves its objective in an
efficient manner.

The Trainee Librarian

Information problems crop up every day. Imagine a trainee librarian on
their first day at work. One thousand brand new books have just been
delivered and are lying in boxes on the floor. The boss wants the books
to be put on the shelves in alphabetical order by author name, as soon
as possible. This is an information problem and there are algorithms for
solving it.
Most people would intuitively use an algorithm called Insertion Sort
(Figure I.1). Insertion Sort operates in the following way:

Figure I.1 Insertion Sort in action.

 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

4 Introduction

Take a pile of unsorted books as input.

Repeat the following steps:
Pick up a book.
Read the author’s name.
Scan across the shelf until you find where the book should
be inserted.
Shift all the books after that point over by one.
Insert the new book.
Stop repeating when there are no books left on the floor.
The books are now sorted.
At any moment in time, the books on the floor are unsorted. One by
one, the books are transferred to the shelf. Every book is placed on the
shelf in alphabetical order. As a result, the books on the shelf are always
in order.
Insertion Sort is easy to understand and works but is slow. It is slow
because, for every book taken from the floor, the librarian has to scan
past or shift every book already on the shelf. At the start, there are
very few books on the shelf, so scanning and shifting is fast. At the end,
our librarian has almost 1,000 books on the shelf. On average, putting a
book in the right place requires 500 operations, where an operation is an
author name comparison or a book shift. Thus, sorting all of the books
takes 500,000 (1,000 × 500) operations, on average. Let’s say that a single
operation takes one second. That being the case, sorting the books using
Insertion Sort will take around seventeen working days. The boss isn’t
going to be happy.
A faster alternative algorithm—Quicksort—was invented by com-
puter scientist Tony Hoare in 1962. Hoare was born in Sri Lanka, to
British parents in 1938. He was educated in England and attended
Oxford University before entering academia as a lecturer. His method
for sorting is a divide-and-conquer algorithm. It is more complicated
than Insertion Sort but, as the name suggests, much faster.
Quicksort (Figure I.2) splits the pile of books into two. The split is
governed by a pivot letter. Books with author names before the pivot letter
are put on a new pile to the left of the current pile. Books with author
names after the pivot are placed on a pile to the right. The resulting
piles are then split using new pivot letters. In doing so, the piles are
kept in sequence. The leftmost pile contains the books that come first
in the alphabet. The next pile holds the books that come second, and
so on. This pile-splitting process is repeated for the largest pile until
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Trainee Librarian 5

Figure I.2 Quicksort in action.

the biggest stack contains just five books. The piles are then sorted
separately using Insertion Sort. Finally, the sorted piles are transferred,
in order, to the shelf.
For maximum speed, the pivot letters should split the piles into two
halves.
Let’s say that the original pile contains books from A to Z. A good
choice for the first pivot would likely be M. This would give two new
piles: A–L and M–Z (Figure I.2). If the A–L pile is larger, it will be split
next. A good pivot for A–L might be F. After this split, there will be
three piles: A–E, F–L, and M–Z. Next, M–Z will be split and so on. For
twenty books, the final piles might be: A–C, D–E, F–L, M–R, and S–Z.
These piles are ordered separately using Insertion Sort and the books
transferred pile-after-pile onto the shelf.
The complete Quicksort algorithm can be written down as follows:

Take a pile of unsorted books as input.

Repeat the following steps:
Select the largest pile.
Clear space for piles on either side.
Choose a pivot letter.
Repeat the following steps:
Take a book from the selected pile.
If the author name is before the pivot letter,
then put the book on the pile to the left,
else put the book on the pile to the right.
Stop repeating when the selected pile is empty.
Stop repeating when the largest pile has five books or less.
Sort the piles separately using Insertion Sort.
Transfer the piles, in order, to the shelf.
The books are now sorted.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

6 Introduction

Quicksort uses two repeating sequences of steps, or loops, one inside

the other. The outer repeating group deals with all of the piles. The
inner group processes a single pile.
Quicksort is much faster than Insertion Sort for large numbers of
books. The trick is that splitting a pile is fast. Each book need only
be compared with the pivot letter. Nothing needs to be done to the
other books—no author name comparisons, no book shifts. Applying
Insertion Sort at the end of Quicksort is efficient since the piles are small.
Quicksort only requires about 10,000 operations to sort 1,000 books. The
exact number of operations depends on how accurately the pivots halve
the piles. At one second per operation, the job takes less than three
hours—a big improvement on seventeen working days. The boss will
be pleased.
Clearly, an algorithm’s speed is important. Algorithms are rated
according to their computational complexity. Computational complexity
relates the number of steps required for execution of an algorithm to
the number of inputs. The computational complexity of Quicksort is
significantly lower than that of Insertion Sort.
Quicksort is called a divide-and-conquer algorithm because it splits
the original large problem into smaller problems, solves these smaller
problems separately, and then assembles the partial solutions to form
the complete solution. As we will see, divide-and-conquer is a powerful
strategy in algorithm design.
Many algorithms have been invented for sorting, including Merge
Sort, Heapsort, Introsort, Timsort, Cubesort, Shell Sort, Bubble Sort,
Binary Tree Sort, Cycle Sort, Library Sort, Patience Sorting, Smooth-
sort, Strand Sort, Tournament Sort, Cocktail Sort, Comb Sort, Gnome
Sort, UnShuffle Sort, Block Sort, and Odd-Even Sort. All of these
algorithms sort data, but each is unique. Some are faster than others.
Some need more storage space than others. A few require that the
inputs are prepared in a special way. A handful have simply been
superseded.
Nowadays, algorithms are inextricably linked with computers. By
definition, a computer is a machine that performs algorithms.

The Algorithm Machine

As discussed, an algorithm is an abstract method for solving a problem.
An algorithm can be performed by a human or a computer. Prior to
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Algorithm Machine 7

execution on a computer, an algorithm must be encoded as a list of

instructions that the computer can carry out. A list of computer
instructions is called a program. The great advantage of a computer
is that it can automatically execute large numbers of instructions
one-after-another at high speed. Surprisingly, a computer need not
support a great variety of instructions. A few basic instruction types will
suffice. All that is needed are instructions for data storage and retrieval,
arithmetic, logic, repetition, and decision-making. Algorithms can be
broken down into simple instructions such as these and executed by a
computer.
The list of instructions to be performed and the data to be operated
on are referred to as the computer software. In a modern computer,
software is encoded as electronic voltage levels on microscopic wires.
The computer hardware—the physical machine—executes the program
one instruction at a time. Program execution causes the input data to
be processed and leads to creation of the output data.
There are two reasons for the phenomenal success of the computer.
First, computers can perform algorithms much more quickly than
humans. A computer can perform billions of operations per second,
whereas a human might do ten. Second, computer hardware is general-
purpose, meaning that it can execute any algorithm. Just change the
software and a computer will perform a completely different task. This
gives the machine great flexibility. A computer can perform a wide
range of duties—everything from word processing to video games. The
key to this flexibility is that the program dictates what the general-
purpose hardware does. Without the software, the hardware is idle. It
is the program that animates the hardware.
The algorithm is the abstract description of what the computer
must do. Thus, in solving a problem, the algorithm is paramount. The
algorithm is the blueprint for what must be done. The program is the
precise, machine-executable formulation of the algorithm. To solve an
information problem, a suitable algorithm must first be found. Only
then can the program be typed into a computer.
The invention of the computer in the mid-twentieth century gave
rise to an explosion in the number, variety, and complexity of al-
gorithms. Problems that were once thought impossible to solve are
now routinely dispatched by cheap computers. New programs are re-
leased on a daily basis, extending the range of tasks that computers can
undertake.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

8 Introduction

Algorithms are embedded in the computers on our desktops, in our

cars, in our television sets, in our washing machines, in our smart-
phones, on our wrists, and, soon, in our bodies. We engage a plethora
of algorithms to communicate with our friends, to accelerate our work,
to play games, and to find our soulmates. Algorithms have undoubtedly
made our lives easier. They have also provided humankind with un-
precedented access to information. From astronomy to particle physics,
algorithms have enhanced our comprehension of the universe. Re-
cently, a handful of cutting-edge algorithms have displayed superhu-
man intelligence.
All of these algorithms are ingenious and elegant creations of the
human mind. This book tells the story of how algorithms emerged from
the obscure writings of ancient scholars to become one of the driving
forces of the modern computerized world.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

1
Ancient Algorithms

Go up on to the wall of Uruk, Ur-shanabi, and walk around,

Inspect the foundation platform and scrutinise the brickwork!
Testify that its bricks are baked bricks,
And that the Seven Counsellors must have laid its foundations!
One square mile is city, one square mile is orchards,
one square mile is clay pits,
as well as the open ground of Ishtar’s temple.
Three square miles and the open ground comprise Uruk.
Unknown author, translated by Stephanie Dalley
The Epic of Gilgamesh, circa 2,000 bce 2

The desert has all but reclaimed Uruk. Its great buildings are almost
entirely buried beneath accretions of sand, their timbers disintegrated.
Here and there, clay brickwork is exposed, stripped bare by the wind or
archaeologists. The abandoned ruins seem irrelevant, forgotten, futile.
There is no indication that seven thousand years ago, this land was the
most important place on Earth. Uruk, in the land of Sumer, was one of
the first cities. It was here, in Sumer, that civilization was born.
Sumer lies in southern Mesopotamia (Figure 1.1). The region is
bounded by the Tigris and Euphrates rivers, which flow from the
mountains of Turkey in the north to the Persian Gulf in the south.
Today, the region straddles the Iran–Iraq border. The climate is hot and
dry, and the land inhospitable, save for the regular flooding of the river
plains. Aided by irrigation, early agriculture blossomed in the ‘land
between the rivers’. The resulting surplus of food allowed civilization
to take hold and flourish.
The kings of Sumer built great cities—Eridu, Uruk, Kish, and Ur.
At its apex, Uruk was home to sixty thousand people. All of life was
there—family and friends, trade and religion, politics and war. We know
this because writing was invented in Sumer around 5,000 years ago.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

10 Ancient Algorithms

Tigris
River

Euphrates
River

Alexandria Babylon
BABYLONIA
Uruk
SUMER Ur
Eridu

Figure 1.1 Map of ancient Mesopotamia and the later port of Alexandria.

Etched in Clay
It seems that writing developed from simple marks impressed on wet
clay tokens. Originally, these tokens were used for record keeping and
exchange. A token might equate to a quantity of gain or a headcount
of livestock. In time, the Sumerians began to inscribe more complex
patterns on larger pieces of clay. Over the course of centuries, simple
pictograms evolved into a fully formed writing system. That system is
now referred to as cuneiform script. The name derives from the script’s
distinctive ‘wedge shaped’ markings, formed by impressing a reed stylus
into wet clay. Symbols consisted of geometric arrangements of wedges.
These inscriptions were preserved by drying the wet tablets in the sun.
Viewed today, the tablets are aesthetically pleasing—the wedges thin
and elegant, the symbols regular, the text neatly organized into rows
and columns.
The invention of writing must have transformed these communities.
The tablets allowed communication over space and time. Letters could
be sent. Deals could be recorded for future reference. Writing facilitated
the smooth operation and expansion of civil society.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Uncovered at Last 11

For a millennium, cuneiform script recorded the Sumerian lan-

guage. In the twenty-fourth century bce, Sumer was invaded by the
armies of the Akkadian Empire. The conquerors adapted the Sumerian
writing methods to the needs of their own language. For a period, both
languages were used on tablets. Gradually, as political power shifted,
Akkadian became the exclusive language of the tablets.
The Akkadian Empire survived for three centuries. Thereafter, the
occupied city states splintered, later coalescing into Assyria in the north
and Babylonia in south. In the eighteenth century bce, Hammurabi,
King of Babylon, reunited the cities of Mesopotamia. The city of Babylon
became the undisputed centre of Mesopotamian culture. Under the
King’s direction, the city expanded to include impressive monuments
and fine temples. Babylonia became a regional superpower. The Akka-
dian language, and its cuneiform script, became the lingua franca of
international diplomacy throughout the Middle East.
After more than one millennium of dominance, Babylon fell, almost
without resistance, to Cyrus the Great, King of Persia. With its capital
in modern Iran, the Persian Empire engulfed the Middle East. Cyrus’s
Empire stretched from the Bosporus strait to central Pakistan and from
the Black Sea to the Persian Gulf. Persian cuneiform script came to
dominate administration. Similar at first glance to the Akkadian tablets,
these new tablets used the Persian language and an entirely different set
of symbols. Use of the older Akkadian script dwindled. Four centuries
after the fall of Babylon, Akkadian fell into disuse. Soon, all understand-
ing of the archaic Sumerian and Akkadian cuneiform symbols was lost.
The ancient cities of Mesopotamia were gradually abandoned. Be-
neath the ruins, thousands of tablets—records of a dead civilization lay
buried. Two millennia passed.

Uncovered at Last
European archaeologists began to investigate the ruins of Mesopotamia
in the nineteenth century. Their excavations probed the ancient sites.
The artefacts they unearthed were shipped back to Europe for inspec-
tion. Amongst their haul lay collections of the inscribed clay tablets.
The tablets bore writing of some sort, but the symbols were now
incomprehensible.
Assyriologists took to the daunting task of deciphering the un-
known inscriptions. Certain oft repeated symbols could be identified
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

12 Ancient Algorithms

and decoded. The names of kings and provinces became clear.

Otherwise, the texts remained impenetrable.
The turning point for translators was the discovery of the Behistun
(Bīsitūn) Inscription. The Inscription consists of text accompanied by a
relief depicting King Darius meting out punishment to handcuffed pris-
oners. Judging by their garb, these captives were from across the Persian
Empire. The relief is carved high on a limestone cliff face overlooking
an ancient roadway in the foothills of the Zagros mountains in western
Iran. The Inscription is an impressive fifteen metres tall and twenty five
metres wide.
The significance of the Inscription only became apparent after Sir
Henry Rawlinson—a British East India Company officer—visited the
site. Rawlinson scaled the cliff and made a copy of the cuneiform text. In
doing so, he spotted two other inscriptions on the cliff. Unfortunately,
these were inaccessible. Undaunted, Rawlinson returned in 1844 and,
with the aid of a local lad, secured impressions of the other texts.
It transpired that the three texts were in different languages—Old
Persian, Elamite, and Babylonian. Crucially, all three recounted the
same propaganda—a history of the King’s claims to power and his mer-
ciless treatment of rebels. Some understanding of Old Persian had per-
sisted down through the centuries. Rawlinson compiled and published
the first complete translation of the Old Persian text two years later.
Taking the Old Persian translation as a reference, Rawlinson and
a loose cadre of enthusiasts succeeded in decoding the Babylonian
text. The breakthrough was the key to unlocking the meaning of the
Akkadian and Sumerian tablets.
The tablets in the museums of Baghdad, London, and Berlin were re-
visited. Symbol by symbol, tablet by tablet, the messages of the Sumeri-
ans, Akkadians, and Babylonians were decoded. A long-lost civilization
was revealed.
The messages on the earliest tablets were simplistic. They recorded
major events, such as the reign of a king or the date of an important
battle. Over time, the topics became more complex. Legends were
discovered, including the earliest written story: The Epic of Gilgamesh.
The day-to-day administration of civil society was revealed—laws, legal
contracts, accounts, and tax ledgers. Letters exchanged by kings and
queens were found, detailing trade deals, proposals of royal marriage,
and threats of war. Personal epistles were uncovered, including love
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Uncovered at Last 13

poems and magical curses. Amid the flotsam and jetsam of daily life,
scholars stumbled upon the algorithms of ancient Mesopotamia.
Many of the extant Mesopotamian algorithms were jotted down by
students learning mathematics. The following example dates from the
Hammurabi dynasty (1,800 to 1,600 bce), a time now known as the Old
Babylonian period. Dates are approximate; they are inferred from the
linguistic style of the text and the symbols employed. This algorithm
was pieced together from fragments held in the British and Berlin State
Museums. Parts of the original are still missing.
The tablet presents an algorithm for calculating the length and
width of an underground water cistern. The presentation is formal and
consistent with other Old Babylonian algorithms. The first three lines
are a concise description of the problem to be solved. The remainder
of the text is an exposition of the algorithm. A worked example is
interwoven with the algorithmic steps to aid comprehension. 5
A cistern.
The height is 3.33, and a volume of 27.78 has been excavated.
The length exceeds the width by 0.83.
You should take the reciprocal of the height, 3.33, obtaining 0.3.
Multiply this by the volume, 27.78, obtaining 8.33.
Take half of 0.83 and square it, obtaining 0.17.
Add 8.33 and you get 8.51.
The square root is 2.92.
Make two copies of this, adding to the one 0.42 and subtracting
from the other.
You find that 3.33 is the length and 2.5 is the width.
This is the procedure.

The question posed is to calculate the length and width of a cistern,

presumably of water. The volume of the cistern is stated, as is its
height. The required difference between the cistern’s length and width
is specified. The actual length and width are to be determined.
The phrase, ‘You should’, indicates that what follows is the method
for solving the problem. The result is followed by the declaration, ‘This
is the procedure’, which signifies the end of the algorithm.
The Old Babylonian algorithm is far from simple. It divides the
volume by the height to obtain the area of the base of the cistern.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

14 Ancient Algorithms

Simply taking the square root of this area would give the length and
width of a square base. An adjustment must be made to create the
desired rectangular base. Since a square has minimum area for a given
perimeter, the desired rectangle must have a slightly larger area than
the square base. The additional area is calculated as the area of a square
with sides equal to half the difference between the desired length and
width. The algorithm adds this additional area to the area of the square
base. The width of a square with this combined area is calculated. The
desired rectangle is formed by stretching this larger square. The lengths
of two opposite sides are increased by half of the desired length–width
difference. The length of the other two sides is decreased by the same
amount. This produces a rectangle with the correct dimensions.
Decimal numbers are used the description above. In the original, the
Babylonians utilized sexagesimal numbers. A sexagesimal number system
possesses sixty unique digits (0–59). In contrast, decimal uses just ten
digits (0–9). In both systems, the weight of a digit is determined by
its position relative to the fractional (or decimal) point. In decimal,
moving right-to-left, each digit is worth ten times the preceding digit.
Thus, we have the units, the tens, the hundreds, the thousands, and
so on. For example, the decimal number 421 is equal to four hundreds
plus two tens plus one unit. In sexagesimal, moving right-to-left from
the fractional point, each digit is worth sixty times the preceding one.
Conversely, moving left-to-right, each column is worth a sixtieth of the
previous one. Thus, sexagesimal 1,3.20, means one sixty plus three units
plus twenty sixtieths, equal to 63 20
60 or 63.333 decimal. Seemingly, the
sole advantage of the Old Babylonian system is that thirds are much
easier to represent than in decimal.
To the modern reader, the Babylonian number system seems bizarre.
However, we use it every day for measuring time. There are sixty
seconds in a minute and sixty minutes in an hour. The time 3:04 am
is 184 (3 × 60 + 4 × 1) minutes after midnight.
Babylonian mathematics contains three other oddities. First, the frac-
tional point wasn’t written down. Babylonian scholars had to infer its
position based on context. This must have been problematic—consider
a price tag with no distinction between dollars and cents! Second, the
Babylonians did not have a symbol for zero. Today, we highlight the
gap left for zero by drawing a ring around it (0). Third, division was
performed by multiplying by the reciprocal of the divisor. In other
words, the Babylonians didn’t divide by two, they multiplied by a half.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Uncovered at Last 15

In practice, students referred to precalculated tables of reciprocals and

multiplications to speed up calculations.
A small round tablet shows the breathtaking extent of Babylonian
mathematics. The tablet—YBC 7289—resides in Yale University’s Old
Babylonian Collection (Figure 1.2). Dating to around 1,800 to 1,600 bce,
it depicts a square with two diagonal lines connecting opposing cor-
ners. The length of the sides of the square are marked as thirty units.
The length of the diagonal is recorded as thirty times the square root
of two.
The values indicate knowledge of the Pythagorean Theorem, which
you may recall from school. It states that, in triangles with a right angle,
the square (a value multiplied by itself) of the length of the hypotenuse
(the longest side) is equal to the sum of the squares of the lengths of the
other two sides.
What is truly remarkable about the tablet is that it was inscribed 1,000
years before the ancient Greek mathematician Pythagoras was born.
For mathematicians, the discovery is akin finding an electric light bulb
in a Viking camp! It raises fundamental questions about the history of
mathematics. Did Pythagoras invent the algorithm, or did he learn of it
during his travels? Was the Theorem forgotten and independently rein-
vented by Pythagoras? What other algorithms did the Mesopotamians
invent?
YBC 7289 states that the square root of two is 1.41421296 (in deci-
mal). This is intriguing. We now know that the square root of two is
1.414213562, to nine decimal places. Remarkably, the value on the tablet

Figure 1.2 Yale Babylonian Collection tablet 7289. (YBC 7289, Courtesy of the Yale
Babylonian Collection.)
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

16 Ancient Algorithms

is accurate to almost seven digits, or 0.0000006. How did the Babylonians

calculate the square root of two so precisely?
Computing the square root of two is not trivial. The simplest method
is Heron of Alexandria’s approximation algorithm. There is, of course,
the slight difficulty that Heron lived 1,500 years (c. 10–70 ce) after
YBE 7289 was inscribed! We must assume that the Babylonians devised
the same method.
Heron’s algorithm reverses the question. Instead of asking ‘What is
the square root of two?’, Heron enquires ‘What number multiplied by
itself gives two?’ Heron’s algorithm starts with a guess and successively
improves it over a number of iterations:

Make a guess for the square root of two.

Repeatedly generate new guesses as follows:
Divide two by the current guess.
Add the current guess.
Divide by two to obtain a new guess.
Stop repeating when the two most recent guesses are almost
equal.
The latest guess is an approximation for the square root of two.

Let’s say that the algorithm begins with the extremely poor guess of:
2.
Dividing 2 by 2 gives 1. Adding 2 to this, and dividing by 2 gives:
1.5.
Dividing 2 by 1.5 gives 1.333. Adding 1.5 to this and dividing by 2 again
gives:
1.416666666.
Repeating once more gives:
1.41421568.
which is close to the true value.
How does the algorithm work? Imagine that you know the true value
for the square root of two. If you divide two by this number, the result
is exactly the same value—the square root of two.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Uncovered at Last 17

Now, imagine that your guess is greater than the square root of two.
When you divide two by this number, you obtain a value less than the
square root of two. These two numbers frame the true square root—
one is too large, the other is too small. An improved estimate can be
obtained by calculating the average of these two numbers (i.e. the sum
divided by two). This gives a value midway between the two framing
numbers.
This procedure—division and averaging—can be repeated to further
refine the estimate. Over successive iterations, the estimates converge
on the true square root.
It is worth noting that the process also works if the guess is less than
the true square root. In this case, the number obtained by means of
division is too large. Again, the two values frame the true square root.
Even today, Heron’s method is used to estimate square roots. An
extended version of the algorithm was utilized by Greg Fee in 1996 to
confirm enumeration of the square root of two to ten million digits.
Mesopotamian mathematicians went so far as to invoke the use of
memory in their algorithms. Their command ‘Keep this number in
your head’ is an antecedent of the data storage instructions available
in a modern computer.
Curiously, Babylonian algorithms do not seem to have contained
explicit decision-making (‘if–then–else’) steps. ‘If–then’ rules were,
however, used by the Babylonians to systematize non-mathematical
knowledge. The Code of Hammurabi, dating from 1754 bce, set out 282
laws by which citizens should live. Every law included a crime and a
punishment: 8
If a son strike a father, they shall cut off his fingers.
If a man destroy the eye of another man, they shall destroy his eye.
If–then constructs were also used to capture medical knowledge and
superstitions. The following omens come from the library of King
Ashurbanipal in Nineveh around 650 bce: 9
If a town is set on a hill, it will not be good for the dweller within that
town.
If a man unwittingly treads on a lizard and kills it, he will prevail over his
adversary.
Despite the dearth of decision-making steps, the Mesopotamians
solved a wide variety of problems by means of algorithms. They
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

18 Ancient Algorithms

projected interest on loans, made astronomical predictions, and even

solved quadratic equations (i.e. equations with unknowns to the power
of two). While most of their algorithms had practical applications, a
few suggest the pursuit of mathematics for its own sake.

Elegance and Beauty

The invention of writing in Egypt using hieroglyphs was roughly con-
temporaneous with its development in Mesopotamia. Due to the use of
perishable papyrus scrolls, little evidence of Egyptian mathematics has
survived to the present day. The most notable extant record is a papyrus
scroll purchased by Henry Rhind in Luxor in 1858. Now in the British
Museum, the Rind Papyrus is an ancient copy of an original dating from
around 2,000 bce. The five-metre-long and thirty-three-centimetre-
wide roll poses a series of problems in arithmetic, algebra, and geometry.
While the basics of these topics are well covered, little of the content is
algorithmic in nature. Overall, it seems that algorithms were not a well-
developed component of ancient Egyptian mathematics.
In the centuries after the rise of the Persian Empire, the Hellenic
world gradually took up the mantle of leadership in mathematics. The
Greeks learned much from the Mesopotamians and Egyptians as by-
products of trade and war.
Alexander the Great (356–323 bce) established Greek military domi-
nance over the entire Middle East in the period 333–323 bce. His con-
quests began with the unification of the Greek city states under his
sole rule by means of military victory. Subsequently, the young man
raised an army of 32,000 infantry allied with 5,000 cavalry and marched
on Asia Minor. Alexander proved to be a brilliant military tactician and
an inspirational leader. His force swept through Syria, Egypt, Phoenicia,
Persia, and Afghanistan, taking city after city. Then, in 323 bce, in the
wake of one of his habitual drinking binges, the Emperor Alexander
was taken with a fever. He died a few days later in Babylon, aged just
32. Alexander’s vast empire was divided between four of his generals.
Ptolemy—a close friend of Alexander’s and, possibly, his half-brother—
was appointed governor of Egypt.
One of Ptolemy’s first decisions was to relocate Egypt’s capital from
Memphis to Alexandria. Alexander himself had founded the city on
the site of an older Egyptian town. Alexandria was ideally located.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Elegance and Beauty 19

Situated on the Mediterranean coast at the western edge of the Nile

delta, its natural harbours afforded the navy and commercial traders
easy access to the Nile. Goods could be transported upriver by barge.
Camel trains linked the Upper Nile with the Red Sea. Alexandria grew
prosperous on trade. Buoyed by an influx of Egyptians, Greeks, and
Jews, Alexandria become the largest city of its time. The historian Strabo
described Alexandria thus: 11
The city also contains very beautiful public parks and royal palaces, which
occupy a fourth or even a third of its whole extent.
There follow along the waterfront a vast succession of docks, military and
mercantile harbours, magazines, also canals reaching the lake Mareôtis,
and many magnificent temples, an amphitheatre, stadium.
In short, the city of Alexandria abounds with public and sacred buildings.
Ptolemy I commissioned construction of the Lighthouse of Alexan-
dria. One of the seven wonders of the ancient world, the mighty light-
house stood on the island of Pharos, which served as a bulwark between
the open sea and the port. Elegant in design, the striking, three-tiered,
100-metre-tall, stone tower supported a beacon—mirror by day and fire
by night—for shipping.
Ptolemy also founded a research institute known as the Mouseion, or
Museum. The ‘Seat of the Muses’ was similar in nature to a modern
research institute, attracting researchers, scientists, authors, and math-
ematicians from around the Mediterranean. Its most famous building
was the renowned Library of Alexandria. The Library was intended to
be a repository for all knowledge. Generously funded, it acquired one
of the largest collections of scrolls in the world. Reputedly, at its zenith,
the library held more than 200,000 books. It is said that all ships entering
the harbour were searched for scrolls. Any material so uncovered was
confiscated and a copy added to the Library. The Library of Alexan-
dria became the preeminent centre of learning in the Mediterranean
world.
Euclid (third–fourth century bce) was perhaps the greatest Alexan-
drian scholar. Little is known of his life save that he opened a school in
the city during the reign of Ptolemy I. Unfortunately, most of Euclid’s
writings are now lost. Copies of five of his books survive. His great work
was Euclid’s Elements of Geometry—a mathematics textbook. It drew on
the writings of his predecessors, running to thirteen chapters covering
geometry, proportions, and number theory. Down through the ages,
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

20 Ancient Algorithms

Euclid’s Elements was copied, translated, recopied, and retranslated. What

is now known as Euclid’s algorithm is contained in Book VII.
Euclid’s algorithm calculates the greatest common divisor of two
numbers (also called the GCD, or the largest common factor). For
example, 12 has six divisors (i.e. integers that divide evenly into it). The
divisors are 12, 6, 4, 3, 2, and 1. The number 18 also happens to have six
divisors: 18, 9, 6, 3, 2, and 1. The greatest common divisor of both 12
and 18 is therefore 6.
The GCD of two numbers can be found by listing all of their divisors
and searching for the largest value common to both lists. This approach
is fine for small numbers, but is very time consuming for large numbers.
Euclid came up with a much faster method for finding the GCD of two
numbers. The method has the advantage of only needing subtraction
operations. Cumbersome divisions and multiplications are avoided.
Euclid’s algorithm operates as follows:

Take a pair of numbers as input.

Repeat the following steps:
Subtract the smaller from the larger.
Replace the larger of the pair with the value obtained.
Stop repeating when the two numbers are equal.
The two numbers are equal to the GCD.

As an example, take the following two inputs:

12, 18.
The difference is 6. This replaces 18, the larger of the pair. The pair is
now:
12, 6.
The difference is again 6. This replaces 12, giving the pair:
6, 6.
Since the numbers are equal, the GCD is 6.
It is not immediately obvious how the algorithm works. Imagine that
you know the GCD at the outset. The two starting numbers must both
be multiples of the GCD since the GCD is a divisor of both. Since both
inputs are multiples of the GCD, the difference between them must
also be a multiple of the GCD. By definition, the difference between the
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Finding Primes 21

inputs has to be smaller than the larger of the two numbers. Replacing
the larger number with the difference means that the pair of numbers
is reduced. In other words, the pair is getting closer to the GCD. At
all times, the pair, and their difference, are multiples of the GCD.
Over several iterations, the difference becomes smaller and smaller.
Eventually, the difference is zero. When this happens, the numbers are
equal to the smallest possible multiple of the GCD, that is, the GCD
times 1. At this point, the algorithm outputs the result and terminates.
This version of Euclid’s algorithm is iterative. In other words, it con-
tains repeating steps. Euclid’s algorithm can, alternatively, be expressed
recursively. Recursion occurs when an algorithm invokes itself. The idea is
that every time the algorithm calls itself, the inputs are simplified. Over
a number of calls, the inputs become simpler and simpler until, finally,
the answer is obvious. Recursion is a powerful construct. The recursive
version of Euclid’s algorithm operates as follows:

Take a pair of numbers as input.

Subtract the smaller from the larger.
Replace the larger with the value obtained.
If the two numbers are equal,
then output one of the numbers – it is the GCD,
else apply this algorithm to the new pair of numbers.

This time, there is no explicit repetition of steps. The algorithm just

calls for execution of itself. Each time, the algorithm is applied to a
smaller pair of numbers: 18 and 12, then 12 and 6, next 6 and 6. Finally,
the inputs are equal, and the result is returned.
The recursive version of Euclid’s algorithm is one of the great algo-
rithms. It is both effective and highly efficient. However, there is more to
it than mere functionality. It has a symmetry, a beauty, and an elegance.
Euclid’s algorithm is an unexpected solution to the problem. It shows
imagination and flair. All of these things make Euclid’s algorithm great.
The great algorithms are poems that solve puzzles.

Finding Primes
In the third century bce, Eratosthenes (c. 276–195 bce) was appointed
Director of the Library of Alexandria. Born in Cyrene, a North African
city founded by the Greeks, Eratosthenes spent most of his early years
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

22 Ancient Algorithms

in Athens. In middle age, he was called by Ptolemy III—grandson of

Ptolemy I—to take charge of the great Library and tutor the King’s son.
Today, Eratosthenes is most famous for having measured the cir-
cumference of the Earth. He discovered that at noon on the summer
solstice (the longest day of the year), the shadow cast by a stake in
the ground at Alexandria is longer than that cast by a stake of equal
height at Syene (now Aswan), 800 km to the south. Measuring the
distance between the cities gave Eratosthenes the length of the arc of
the Earth between Alexandria and Syene. Combining this with the ratio
of the shadow lengths, he produced an estimate the circumference of
the Earth. Amazingly, his calculation of five times the distance between
the two cities was accurate to within sixteen per cent of the true value.
As part of his researches on mathematics, Eratosthenes invented
an important algorithm for finding prime numbers—the Sieve of Er-
atosthenes. A prime number has no exact whole number divisors (i.e.
numbers that divide into it evenly) other than itself and one. The first
five primes are 2, 3, 5, 7, and 11.
Primes are notoriously difficult to find. There are infinitely many,
but they are scattered randomly across the number line. Even with
modern computers, discovering new primes is time consuming. Some
algorithms provide shortcuts but, to date, there is no easy way to find
all primes.
The Sieve of Eratosthenes operates as follows:

List the numbers that you wish to find primes among, starting
at 2.
Repeat the following steps:
Find the first number that hasn’t been circled or crossed out.
Circle it.
Cross out all multiples of this number.
Stop repeating when all the numbers are either circled or
crossed out.
The circled numbers are prime.

Imagine trying to find all of the primes up to fifteen. The first step is
to write down the numbers from 2 to 15. Next, circle 2 and cross out its
multiples: 4, 6, 8, and so on.
2 , 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Finding Primes 23

Then, circle 3 and cross out all of its multiples: 6, 9, 12, 15.
2 , 3 , 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
Four is already crossed out, so the next number to circle is 5, and so it
goes. The final list is:
2 , 3 , 4, 5 , 6, 7 , 8, 9, 10, 11 , 12, 13 , 14, 15
The numbers that pass through the sieve (that is, are circled) are prime.
One of the neat aspects of the Sieve of Eratosthenes is that it does not
use multiplication. Since the multiples are generated sequentially, one
after another, they can be produced by repeatedly adding the circled
number to a running total. For example, the multiples of 2 can be
calculated by repeatedly adding two to a running total, giving 4, 6, 8,
and so on.
A drawback of the Sieve is the amount of storage that it needs. To
produce the first seven primes, eighteen numbers have to be stored. The
amount of storage can be reduced by only recording whether a number
has been struck off or not. Nevertheless, the storage complexity of the Sieve
becomes a problem for large numbers of primes. An up-to-date laptop
computer can find all primes with fewer than eight decimal digits using
the Sieve of Eratosthenes. In contrast, as of March 2018, the largest
known prime number contains a whopping 23,249,425 decimal digits.
For three hundred years, the Museum of Alexandria was a beacon
of teaching and learning. Thereafter, slow decline was punctuated by
disaster. In 48 bce, Julius Caesar’s army put their vessels to the torch
in Alexandria’s harbour in a desperate attempt to stall Ptolemy XIV’s
forces. The fire spread to the docks and parts of the Library were
damaged in the resulting conflagration. The Museum was damaged in
an Egyptian revolt in 272 ce. The Temple of Serapis was demolished in
391 ce by the order of Coptic Christian Pope Theophilus of Alexandria.
The female mathematician Hypatia was murdered in 415 ce by a Chris-
tian mob. The Library was finally destroyed when General Amr ibn
al- As.al–Sahmī’s army took control of the city in 641 ce.
While the Museum of Alexandria was the pre-eminent centre of
learning in the ancient Greek world for six centuries, it was not the only
stronghold of logic and reason. On the other side of the Mediterranean,
a lone genius invented a clever algorithm for calculating one of the
most important numbers in all of mathematics. His algorithm was to
outshine all others for nearly a thousand years.
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

2
Ever-Expanding Circles

How I wish I could recollect of circle round,

The exact relation Archimede unwound.
Unknown author, recounted by J. S. MacKay
Mnemonics for π , π1 , e, 1884 13

Göbekli Tepe lies in southern Turkey, close to the headwaters of the

Euphrates river. Excavations at the site have uncovered a series of
mysterious megalithic structures. Four-metre-high limestone pillars
are set out in ten- to twenty-metre-wide circles. The circles are centred
on pairs of still larger monoliths. The pillars are shaped similarly to an
elongated ‘T’. Most are richly engraved with depictions of animals. In
places, the motifs are reminiscent of human hands and arms. In total,
there are twenty circles and about 200 pillars.
While the structures are impressive, the truly extraordinary aspect
of Göbekli Tepe is its age. The site dates from 10,000–8,000 bce, far
pre-dating ancient Sumer. This makes Göbekli Tepe the oldest known
megalithic site in the world.
The practice of megalithic circle building was still in existence in
Europe more than six thousand years later. One wonders what is so
special about the circle that humankind chose to incorporate it in its
greatest monuments over such a long period?

Wheels Within Wheels

The fundamental characteristic of the circle is that the distance from
its centre to its perimeter is constant. This distance is the circle’s radius.
A circle’s diameter, or width, is twice its radius. The circumference of a circle
is the length of its perimeter. The larger a circle, the greater both its
circumference and diameter. An assessment of the relationship between
circumference and diameter can be made by measurement. Stretch a
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

26 Ever-Expanding Circles

piece of string across a circle’s diameter and compare that length to the
circumference. You will find that the circle’s circumference is slightly
more than three times its diameter. Repeated measurement shows that
this ratio is constant for all circle sizes. Of course, ‘slightly more than
three times’ isn’t particularly satisfactory from a mathematical point
of view. Mathematicians want precise answers. Determining the precise
ratio of a circle’s circumference to its diameter is a never-ending quest.
The exact ratio—whatever its true value—is today represented by
the Greek letter π (pronounced ‘pi’). The letter π was first used in this
way, not by the ancient Greeks, but by a Welshman—mathematician
William Jones in 1707.
Writing down the true value of π is impossible. Johann Heinrich
Lambert proved that π is an irrational number, meaning that enumeration
requires infinitely many digits (1760s). No matter how far enumeration
goes, the digits never settle into a repeating pattern. The best anyone
can do is approximate π.
After the first few integers, π is arguably the most important number
in mathematics. Without π, we would struggle to reason about circles
and spheres. Circular motion, rotations, and vibrations would become
mathematical conundrums. The value of π is employed in many prac-
tical applications ranging from construction to communication, and
from spaceflight to quantum mechanics.
The original estimate of 3 is correct to 1 digit. By about 2,000 bce,
the Babylonians had estimated π as 258 = 3.125, accurate to two digits.
The Egyptian Rhind Papyrus offers an improved approximation of
81 = 3.16049, close to three-digit accuracy. However, the first real
256

breakthrough in determining π came from the Greek mathematician

Archimedes.
Archimedes (c. 287–212 bce) is considered to have been the greatest
mathematician of antiquity. He was born in the city of Syracuse, Sicily,
then part of a Greek colony.
The details of Archimedes’ life are mostly unknown. Today, he is
remembered for having leapt from his bathtub to run down the street
yelling ‘Eureka!’ (‘I have found it!’). That tale comes from the histories
of Vitruvius. It seems that Archimedes had been asked by the king to
inspect the royal crown. The king was suspicious that his goldsmith had
surreptitiously substituted a cheap silver–gold alloy for pure gold. The
alloy looked identical to the real thing. Could Archimedes determine
the truth?
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

Wheels Within Wheels 27

There is one measurable difference between a silver–gold alloy and

pure gold: pure gold is more dense. The density of an object is its weight
(or mass) divided by its volume. The weight of the crown could be
measured. However, determining its volume seemed impossible, as its
shape was irregular.
One fateful night, Archimedes’ bath happened to overflow as he
climbed in. In a flash, Archimedes realised that the volume of an irreg-
ular object can be determined by measuring the quantity of water that
it displaces when immersed. This deduction allowed him to determine
the density of the crown.
The crown was not pure gold. The goldsmith was guilty.
Archimedes solved a series of important problems in mechanics,
including defining the law of the lever. His greatest contributions,
however, were in geometry. His studies led to him to enquire upon
the correct value for π. The result of his labours was an algorithm for
calculating π with unheard-of precision.
Archimedes’ algorithm for approximating π is based on three in-
sights. Firstly, a regular polygon approximates a circle. Secondly, it
is easy to calculate the perimeter of a polygon because its sides are
straight. Thirdly, the more sides a regular polygon has, the closer its
approximation to a circle.
Imagine a circle. Now, visualize a hexagon (a regular six-sided figure)
drawn just inside the circle (Figure 2.1). The corners of the hexagon
touch the circle’s perimeter and its sides lie just inside the circle. Since
the hexagon is smaller than the circle, it stands to reason that the
perimeter of the hexagon is close to, but slightly less than, that of the
circle.
A regular hexagon is equivalent in outline to six identical triangles
placed edge-to-edge pointing towards the centre. These triangles are
equilateral, that is, all three sides are the same length. Since a hexagon
has six edges, its perimeter is equal to six triangle sides. The diameter of
a hexagon is equal to two triangle sides. Thus, the ratio of a hexagon’s
perimeter to its diameter is 62 = 3. Hence, 3 is a reasonable approxima-
tion for π.
Now, consider a hexagon drawn just outside the circle (Figure 2.1).
In this case, the middle of each hexagon side touches the circle, not the
corners. The diameter of the circle is now equal to twice the distance
from the centre of the polygon √ to the middle of one side. The perimeter
of this, larger, hexagon is 2 3 times the diameter of the circle. This gives
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

28 Ever-Expanding Circles

Figure 2.1 A circle with an inner hexagon (left) and a circle with an
outer hexagon (right). The inner hexagon includes its constituent equilateral
triangles.

another estimate for π equal to 3.46410. This estimate is close to the true
value but is a little too large.
Archimedes improved these approximations by means of an algo-
rithm. Every iteration of the algorithm doubles the number of sides
in the two polygons. The more sides a polygon has, the better its
approximation to π.
The algorithm operates as follows:

Take the perimeters of a pair of inner and outer polygons as

input.
Multiply the inner and outer perimeters.
Divide by their sum.
This gives the perimeter of a new outer polygon.
Multiply this new outer perimeter by the previous inner
perimeter.
Take the square root.
This gives the perimeter of a new inner polygon.
Output the perimeters of the new inner and outer polygons.

In the first iteration, the algorithm turns the hexagons into do-
decagons (12-sided figures). This gives improved estimates for π of
3.10582 (inner polygonal) and 3.21539 (outer polygon) to six digits.
The beauty of Archimedes’ algorithm is that it can be applied again.
The outputs from one run can be fed into the algorithm as inputs to
the next iteration. In this way, the dodecagons can be transformed
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

World Records 29

into twenty-four-sided polygons. Forty-eight-sided polygons can be

turned into ninety-six-sides figures, and so on. With every repetition,
the inner and outer polygons close in on the circle, providing better
estimates for π.
Archimedes completed the calculations for a ninety-six-sided figure,
obtaining estimates for π of 223 22
71 and 7 . The former is accurate to four
digits. The latter is less accurate but more popular due to its simplicity.
Tragically, Archimedes was slain by a Roman soldier during the
sack of Syracuse. Accounts differ as to the provocation. In one telling,
Archimedes declined to accompany the soldier to his superior officer on
the grounds that he was working on a particularly intriguing problem.
In another, Archimedes attempted to prevent the soldier stealing his
scientific instruments. Amazingly, it would be nearly two thousand
years before Archimedes’ algorithm was surpassed (1699 ce).

World Records
Archaeological evidence suggests that civilization emerged in China at
around the same time as it appeared in Mesopotamia and Egypt. Urban
society in China appears to have first developed along the banks of the
Yangtze and Yellow Rivers. Little is known of early Chinese mathemat-
ics as the bamboo strips used for writing at the time were perishable.
Although there was intercommunication between East and West, it
appears that Chinese mathematics developed largely independently.
The oldest extant Chinese mathematical text—Zhoubi Suanjing—dates
to around 300 bce. The book focuses on the calendar and geometry, and
includes the Pythagorean Theorem. A compendium of mathematical
problems analogous to the Rhind Papyrus—Nine Chapters on the Mathe-
matical Art—survives from roughly the same period.
It seems that the search for π was much more determined in China
than in the West. In 264 ce, Liu Hui used a ninety-six-sided inner
polygon to obtain an approximation of 3.14—accurate to three figures.
He later extended his method to a polygon of 3,072 sides, obtaining an
improved estimate of 3.14159—accurate to six figures.
Zu Chongzhi (430–501 ce), assisted by his son Zu Chengzhi, pro-
duced an even better estimate in the fifth century ad. The father
and son duo used a polygonal method similar to Archimedes’ ap-
proach. However, they persevered for many more iterations. Their up-
per and lower bounds of 3.1415927 and 3.1415926 were accurate to seven
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

30 Ever-Expanding Circles

digits—a new world record. Their accomplishment was to stand for

nigh on nine hundred years, a testament to their dedication
Today, we know that π is equal to 3.14159265359 to twelve figures.
Enumeration of π is now an endeavour for computer algorithms.
According to the Guinness Book of World Records, the most accurate
value of π is thirty-one trillion digits long. The programme that pro-
duced the value was written by Emma Haruka Iwao of Japan. Running
on twenty-five virtual machines in the Google Cloud, the calculation
took 121 days.

The Art of Reckoning

Archimedes’ murder in 212 bce was a harbinger of the Roman domi-
nation of Europe. Ancient Greece fell to the might of Rome in 146 bce.
From the first century bce to the 5th century ce, the Roman Empire
controlled the Mediterranean. When the Empire finally fell, European
civilization disintegrated. The flame of European mathematics flickered
and guttered for a thousand years. Amidst the darkness, a few centres
of learning illuminated the East.
Caliph Harun al-Rashid founded the House of Wisdom (Bayt al-
Hikmah) in his new capital of Baghdad around 762 ce. Expanded by his
successors, the House grew to become a major intellectual centre from
the ninth to the thirteenth century, a period now known as the Islamic
Golden Age. Scholars working in the House translated scientific and
philosophical texts written in Greek and Indian into Arabic. They also
conducted original research in mathematics, geography, astronomy,
and physics.
The House of Wisdom’s most influential intellectual was Muhammad
ibn Mūsā al-Khwārizmī. Al-Khwārizmī lived in Baghdad from around
780 to 850. Not much is known of his life. However, copies of three of
his major works survive.
The Compendious Book on Calculation by Completion and Balancing focuses on
algebra. In fact, we derive the English word ‘algebra’ from the book’s
Arabic title (al-jabr, meaning completion). The book describes how to
use algorithms to solve mathematical problems, especially linear and
quadratic equations. While algebra had been described previously, it was
al-Khwārizmī’s style of presentation that caught the eye. His treatment
was more systematic, more step-by-step, more algorithmic, than that
found in other works.
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

Waves Upon Waves 31

al-Khwārizmī’s text On the Hindu Art of Reckoning (c. 825) describes the
decimal number system, including the numerals that we employ today.
The system’s roots lie in the Indus Valley civilization (now southern
Pakistan), which blossomed around 2,600 bce—roughly contempora-
neous with the construction of the pyramids in Giza. Little is known
of the original mathematics of the region, save what can be gleaned
from religious texts. Inscriptions suggest that the nine non-zero Hindu–
Arabic numerals (1–9) appeared in the region between the third century
bce and the second century ce. Certainly, there is a clear reference to the
nine non-zero Hindu numerals in a letter from Bishop Severus Sebokht,
who lived in Mesopotamia around 650 ce. The numeral for zero (0)
finally appeared in India at around the same time.
By the eighth century, many Persian scholars had adopted the
Hindu–Arabic number system by reason of its great convenience—
hence, al-Khwārizmī’s book on the subject. His text became a conduit
for the Hindu–Arabic numerals in their transfer to the West. On the
Hindu Art of Reckoning was translated in 1126 from Arabic into Latin by
Adelard of Bath–an English natural philosopher. Adelard’s translation
was followed by Leonardo of Pisa’s (Fibonacci) book on the topic—Liber
Abaci—in 1202.
In 1258, four hundred years after al-Khwārizmī’s death, The House of
Wisdom was destroyed in the Mongol sack of Baghdad.
Surprisingly, uptake of the new number system was slow. It would be
centuries before Roman numerals (I, II, III, IV, V, …) were displaced by
the Hindu–Arabic digits. It seems that European scholars were perfectly
happy to perform calculations on an abacus and record the results
using Roman numerals. Decimal numbers only became the preferred
option in the sixteenth century with the transition to pen and paper
calculation.
It is al-Khwārizmī’s name in the title of a Latin translation of his
book—Algoritmi de Numero Indorum—that gives us the English word
‘algorithm’.

Waves Upon Waves

The European Renaissance of the fourteenth to seventeenth centuries
saw the rediscovery of classical philosophy, literature, and art. Mathe-
matics, too, was resurgent, particularly in its application to practicalities
such as accountancy, mechanics, and map-making. The invention of
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

32 Ever-Expanding Circles

the printing press in the fifteenth century further spurred the spread
of scholarship and learning.
The ensuing Enlightenment of the eighteenth century saw a revolu-
tion in Western philosophy. Centuries of dogma were swept away by the
strictures of evidence and reason. Mathematics and science became the
foundations of thought. Technological progress altered the very fabric
of society. Democracy and the pursuit of individual liberty were in the
ascendency.
Changing attitudes, heavy taxation, and failed harvests were to ignite
the French Revolution in 1789. Amidst the bloody upheaval, a French
mathematician laid the theoretical foundations for what was to become
one of the world’s most frequently used algorithms.
In 1768, Jean-Baptiste Joseph Fourier (Figure 2.2) was born in Auxerre,
France. Orphaned at age nine, Fourier was educated in local schools
run by religious orders. The lad’s talent for mathematics became ob-
vious as he entered his teenage years. Nonetheless, the young man
undertook to train for the priesthood. On reaching adulthood, Fourier
abandoned the ministry to devote his career to mathematics, taking
up employment as a teacher. Soon, he became entangled in the po-
litical upheaval sweeping the country. Inspired by the ideals of the

Figure 2.2 Bust of Joseph Fourier by Pierre-Alphonse Fessard, 1839.

(© Guillaume Piolle / CC-BY 3.0.)
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

Waves Upon Waves 33

Revolution, Fourier turned to political activism and joined the Auxerre

Revolutionary Committee. In the Terror that followed, Fourier found
himself implicated in a violent dispute between rival factions. He was
imprisoned and only narrowly escaped the guillotine.
Subsequently, Fourier moved to Paris to study as a teacher. His
mathematical aptitude led him to be appointed as a member of faculty
at the newly established École Polytechnique. A mere two years later,
he was Chair of Analysis and Mechanics. Fourier seemed destined for a
life in academia until an unexpected turn of events changed the course
of his life.
Fourier was appointed scientific advisor to Napoléon’s army for the
invasion of Egypt. The French forces took Alexandria on 1 July 1798.
Thereafter, victory turned to ignominious defeat. Napoléon departed
Egypt, but Fourier stayed on in Cairo. There, he dispensed his scientific
duties and, in his spare time, conducted research on the antiquities of
Egypt.
On Fourier’s eventual repatriation, Napoléon arranged for the math-
ematician to be appointed Prefect (administrator) of the Department of
Isère—the Alpine region centred on the city of Grenoble. It was here
that Fourier embarked on his magnum opus. His Théorie Analytique de la
Chaleur was published in 1822. Nominally it was about the conduction
of heat in metal bars. More importantly, the book suggested that any
waveform can be viewed as the summation of appropriately delayed
and scaled harmonic waves. The idea was highly controversial at the
time. Nevertheless, Fourier’s postulate was subsequently proven to be
correct.
To understand Fourier’s proposal, we will conduct a thought exper-
iment. Imagine a swimming pool. Put a wave machine at one end and
assume that the opposite end has some sort of overflow that doesn’t
reflect waves. Let us say that the machine produces a single wave that
is the length of the pool. We watch the crest move from one end of the
pool to the other before the next crest appears back at the start. This
simple waveform is called the first harmonic (Figure 2.3). Its period—the
distance between two crests—is equal to the length of the pool.
Now, image that the wave machine moves faster. This time, two
complete cycles of the waveform fit into the length of the pool. In other
words, we can always see two crests, not just one. This waveform is the
second harmonic. Its period is equal to half the length of the pool.
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

34 Ever-Expanding Circles

Figure 2.3 Three harmonics (left) and the waveform resulting from their
summation (right). The second harmonic is scaled by a half and the third
harmonic is delayed by a quarter cycle.

Next, we double the speed of the machine again. This time, the period
is a quarter of the length of the pool. This is the third harmonic.
Another doubling and we get the fourth harmonic, and so on.
This sequence of harmonics is called the Fourier series.
Fourier’s remarkable idea was that all waveforms—of any shape
whatsoever—are the sums of scaled and delayed harmonics. Scaling
means increasing or decreasing the size of the waveform. Scaling up
makes the peaks higher and the troughs lower. Scaling down does the
converse. Delaying a waveform means shifting it in time. A delay means
that the wave’s crests and troughs arrive later than before.
Let us examine the effect of combining harmonics (Figure 2.3). Let
us say that the first harmonic has an amplitude of one. The amplitude
of a waveform is its maximum deviation from rest. The amplitude of
a harmonic is the height of the crests. The second harmonic has an
amplitude of a half. The third harmonic has an amplitude of one and
a delay of half a period. If we add these harmonics up, we get a new
compound waveform. The process of addition mimics what happens in the
real world when waves meet. They simply ride on top of one another.
The terminology used in physics is to say that the waves superpose.
Clearly, adding waveforms is easy. Reversing the process is much
more complicated. How, given a compound waveform, can the am-
plitudes and delays of the constituent harmonics be determined? The
answer is by means of the Fourier transform (FT).
The FT takes any waveform as input and breaks it up into its com-
ponent harmonics. Its output is the amplitude and delay of every
harmonic in the original waveform. For example, given the compound
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

Waves Upon Waves 35

waveform in Figure 2.3, the FT will output the amplitudes and delays of
the three constituent harmonics. The output of the FT is two sequences,
one of which lists the amplitudes of the harmonics. For the compound
waveform in Figure 2.3, the amplitudes of the constituent harmonics
are [1, 12 , 1]. The first entry is the amplitude of the first harmonic, and so
on. The second output sequence is the delays of the harmonics: [0, 0, 12 ]
measured in periods.
While initially only of interest to physicists, the real power of the
FT became evident in the decades after the invention of the computer.
Computers allow all kinds of waveforms to be quickly and cheaply
analysed.
A computer stores a waveform as a list of numbers (Figure 2.4). Every
number indicates the level of the waveform at a particular moment
in time. Large positive values are associated with the peaks of the
waveform. Large negative numbers with the troughs. These numbers
are called samples since the computer takes a ‘sample’ of the level the
waveform at regular intervals of time. If the samples are taken often
enough, the list of numbers gives a reasonable approximation to the
shape of the waveform.
The FT commences by generating the first harmonic. The generated
waveform contains a single period—one crest and one trough—and is
the same length as the input sequence. The algorithm multiplies the
two lists of numbers sample-by-sample (e.g. [1, −2, 3] × [10, 11, 12] =
[10, −22, 36]). These results are then totalled. The total is the correlation
of the input and the first harmonic. The correlation is a measure of the
similarity of two waveforms. A high correlation indicates that the first
harmonic is strong in the input.
The algorithm repeats the correlation procedure for a copy of the
first harmonic delayed by a quarter of a period. This time, the corre-

Figure 2.4 A waveform and the associated sample values that are used to
represent the signal.
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

36 Ever-Expanding Circles

lation measures the similarity between the input waveform and the
delayed first harmonic.
The two correlation values (not delayed and delayed) are fused to
produce an estimate of the amplitude and delay of the first harmonic.
The amplitude is equal to the square root of the sum of the squares of
the two correlations divided by the number of samples. The delay is
obtained by calculating the relative strength of the two correlations.
The relative strength indicates how close the component is in time to
the two versions of the harmonic.
This double correlation (not delayed and delayed) and fusion process
is repeated for all higher harmonics. This gives their amplitudes and
delays.
In summary, the FT algorithm works as follows:

Take a waveform as input.

Repeat the following steps for every possible harmonic:
Generate the harmonic.
Correlate the input and harmonic.
Delay the harmonic by a quarter period.
Correlate the input and delayed harmonic.
Calculate the overall amplitude and delay of the harmonic.
Stop repeating when all harmonics have been processed.
Output the amplitude and delay of every harmonic.

On first inspection, this process of breaking a compound waveform

into its constituent harmonics sounds like a mathematical conceit—a
nice thing to demonstrate but of little practical use. This view could not
be further from the truth! The FT is used extensively in the analysis of
real-world signals.
A signal is any real-world quantity that varies with time. Sound sig-
nals are variations in air pressure that we can hear. In speech recognition
systems (e.g. Siri, Alexa), the FT breaks the sound signal up into its
constituent harmonics prior to further analysis. Digital music players
(e.g. Spotify, Tidal), rely on the FT to identify redundant harmonic
information and so reduce data storage requirements.
Radio signals are electromagnetic variations that can be detected
using electronic equipment. In wireless communication systems
(e.g. WiFi, DAB), the FT allows data to be efficiently transmitted and
received by means of radio signals.
 OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi

Waves Upon Waves 37

While the FT is extremely useful, it requires a lot of computing

power. The correlations take an inordinate amount of time, particularly
for long waveforms. A variant of the original algorithm—appropriately
named the fast Fourier transform, or FFT—is used in modern devices.
The FFT was invented in 1965 to meet a pressing military need.
With the Cold War on the horizon, the US wished to monitor Soviet
nuclear tests. The only way to do so was to measure blast induced
seismic vibrations at monitoring stations sited in friendly countries.
Unfortunately, analysis of the seismic data using the traditional FT was
prohibitively slow. Scientists James Cooley (1926–2016) and John Tukey
(1915–2000) conjured up a new version of the algorithm that produced
the same results in a fraction of the time.
Their algorithm exploits the symmetries in harmonic waveforms to
share results between correlation stages. In a harmonic waveform, a
trough is simply a reflection of a peak. The rising half of a crest is a
mirror image of the falling part. Higher harmonics are accelerated ver-
sion of lower harmonics. By re-using intermediate results, unnecessary
repetitions in the calculation can be eliminated.
The idea worked a treat. The Cooley–Tukey FFT allowed the US
military to locate Soviet nuclear tests to within fifteen kilometres.
Unexpectedly, almost twenty years after the ‘invention’ of the FFT,
it was revealed that the algorithm was, in fact, more than 180 years
old. The great German mathematician Carl Friedrich Gauss (1777–1855)
had employed the algorithm in the analysis of astronomical data in
1805. Ever the perfectionist, Gauss never got around to publishing it.
The method was finally uncovered amidst a posthumous collection of
Gauss’s papers. Gauss’s 1805 notes even pre-date Fourier’s work on the
topic. In retrospect, it seems that the algorithm might more properly
have been called the Gauss transform!
Fourier died on 16 May 1830. His name was later engraved on the side
of the Eiffel Tower in recognition of his scientific achievements.
Fourier’s turbulent life coincided almost exactly with the years of
the Industrial Revolution. Over the course of a mere seventy years,
ancient handcrafts were supplanted by machine production. Amidst
the changes, an Englishman got to wondering whether a machine could
weave calculations, rather than cloth. Might there be an Industrial
Revolution for arithmetic?
OUP CORRECTED PROOF – FINAL, 14/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

3
Computer Dreams

The most obvious and the most distinctive features of the History
of Civilization, during the last fifty years, is the wonderful increase
of industrial production by the application of machinery, the
improvement of old technical processes and the invention of
new ones.
Thomas Henry Huxley
The Advance of Science in the Last Half-Century, 1887 25

Performing calculations manually is slow and tedious. For millennia,

inventors have sought to design devices that accelerate arithmetic. The
first success was the abacus. Invented by the Sumerians around 2,500 bce,
the table abacus evolved from counting with pebbles and sand writing.
Later, lines and symbols were etched into the table surface to facilitate
more rapid computation. The bead-and-rail abacus was invented in
China before being popularized in Europe by the ancient Romans.
The first mechanical calculators were invented independently in
France and Germany in the seventeenth century. Driven by hand
cranks, additions and subtractions were performed by the movement
of levers, cogs, and gears. The intricate hand-crafted devices were both
expensive and unreliable. Most were sold to the wealthy as mere curios.
The end of the eighteenth and the beginning of the nineteenth
century brought the Industrial Revolution. Engineers devised machines
powered by steam and running water that replaced traditional manual
production methods. The transition to machine production led to rapid
increases in productivity and major societal changes. Labourers were
displaced from the countryside to the towns and cities to work alongside
the machines in noisy, dark, and often dangerous factories.
Powered textile looms produced fabric with a fixed weave. In 1804,
Joseph Marie Charles (‘Jacquard’)—a French weaver and merchant—
came up with a radical redesign. Jacquard’s loom could be programmed
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

40 Computer Dreams

to produce cloth with different weaves. The machine wove fabric ac-
cording to the pattern of holes cut into cards. By altering the card
punch pattern, a different weave could be produced. Sequences of cards
were linked together in loops such that the machine could repeat the
programmed pattern.
Thus, by the early nineteenth century, Europe was in possession of
hand-powered calculators and steam-powered programmable looms.
An English mathematician with a childhood fascination for machines
began to wonder about combining these concepts. Surely, a steam-
powered programmable machine could perform calculations much
faster, and more reliably, than any human? His idea almost changed
the world.

A Clockwork Computer
Charles Babbage was born in England in Walworth, Surrey in 1791
(Figure 3.1). The son of a wealthy banker, Babbage became a student

Figure 3.1 Charles Babbage (left, c. 1871) and Ada Lovelace (right, 1836),
the first programmers. (Left: Retrieved from the Library of Congress, www.loc.gov/item/
2003680395/. Right: Government Art Collection. GAC 2172, Margaret Sarah Carpenter:
(Augusta) Ada King, Countess of Lovelace (1815–1852) Mathematician; Daughter of Lord
Byron.)
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

A Clockwork Computer 41

of mathematics. Self-taught initially, he was admitted to Cambridge

University at eighteen. Once there, he found the Mathematics depart-
ment to be rather staid and disappointing. A headstrong young man,
Babbage did not bother to ingratiate himself with either his examiners
or his prospective employers. Despite being a superb mathematician,
he failed to find a position in academia upon graduation. Supported
by an allowance from his father, Babbage resolved to conduct his own
independent mathematical research. He moved to London and entered
into the scientific life of the capital, publishing a series of well-regarded
papers.
Scientific papers, like those written by Babbage, are the lifeblood of
research. A paper is a report describing a new idea backed up with
supporting experimental results or a mathematical proof, ideally both.
Science relies on evidence. Proofs must be verified. Experiments must
be repeatable. Papers are reviewed by experts in the field. Only the best
are accepted for publication. Crucially, the idea presented in a paper
must be novel and proven. Publication in journals, or at conferences,
propagates new ideas to interested parties in the scientific community.
Publication of a paper is a milestone in an academic’s career—an indi-
cation of their prowess and standing in the field.
Lacking public funding, science was conducted in a handful of
universities and by a small number of wealthy enthusiasts. Scientific
discourse often took place in the salons of the rich. Even the word
‘scientist’ was new. For years, Babbage was the quintessential Victorian
gentleman scientist. He was eventually appointed Lucasian Professor
of Mathematics at Cambridge in 1828. His innate talents were greatly
amplified by his capacity for long, arduous working hours. Although an
accomplished mathematician, perhaps his primary gift lay in invention
of mechanical devices.
By dint of his published contributions, Babbage was elected a Fellow
of the Royal Society. One of the duties that befell him was reviewing
mathematical tables for the Astronomical Society. The tables listed the
predicted times and positions of notable celestial events. These tables
were used extensively by seafarers as an aid to navigation. Laborious to
produce by means of manual calculation, the tables often contained
errors. On the high seas, an error could lead to shipwreck.
To alleviate the workload in their production, Babbage produced
a design for a steam-powered mechanical machine that could auto-
matically calculate and print the tables. Decimal numbers were to be
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

42 Computer Dreams

represented by the positions of gears, cogs, and levers. The engine would
be capable of automatically performing a sequence of calculations,
storing, and re-using intermediate results along the way. The machine
was designed to perform a single, fixed algorithm. Hence, it lacked
programmability. Nevertheless, the design was a significant advance on
previous calculators, which required manual entry of each and every
number and operation. Babbage fabricated a small working model of
the machine. Seeing merit in Babbage’s concept, the British government
agreed to fund construction of Babbage’s Difference Engine.
Building the Engine proved challenging. Tiny inaccuracies in fab-
rication of its components made the machine unreliable. Despite re-
peated investment by the government, Babbage and his assistant Jack
Clement only completed a portion of the machine before construction
was abandoned. In total, the British Treasury spent nearly £17,500 on
the project. No small sum, the amount was sufficient to otherwise
procure twenty-two brand new railway locomotives from Mr. Robert
Stephenson.26
Despite the failure of his Difference Engine project, Babbage was still
drawn to the idea of automated calculation. He designed a new, much
more advanced machine. The Analytic Engine was to be mechanical,
steam-driven, and decimal. It would also be fully programmable. Bor-
rowing from Jacquard’s loom, the new machine would read instruc-
tions and data from punch cards. Likewise, results would be proffered on
punched cards. The Analytic Engine was to be the first general-purpose
computer.
Once again, Babbage appealed to the government for funding. This
time, the money was not forthcoming. The Analytic Engine project
stalled.
Babbage made his only public presentation on the Analytic Engine to
a group of mathematicians and engineers in Torino, Italy. One of the
attendees—Luigi Federico Menabrea—a military engineer, made notes
and, subsequently, with the help of Babbage, published a paper on the
device. That paper was in French. Another supporter of Babbage’s—
Ada Lovelace—greatly admired the work and resolved to translate it
into English.
Ada Lovelace (born Augusta Ada Byron) was born in 1815, the
daughter of Lord and Lady Byron (Figure 3.1). Lady Byron (Anne
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

A Clockwork Computer 43

Isabella Noel, née Milbanke) was herself a mathematician. Her husband,

Lord Byron, is still regarded as one of the great English poets. Lady and
Lord Byron’s marriage only lasted a year before the pair separated. Lady
Byron told of her husband’s dark moods and her mistreatment. 26
Rumours persisted of his infidelity. Disgraced, the poet left England,
never to see his daughter again.
At her mother’s bidding, Lovelace grew up studying science and
mathematics. These were sensible subjects, far from the worrisome in-
fluence of poetry and literature. Aged only seventeen, she met Babbage
in 1833 at one of his social gatherings. At the time, Babbage was a forty-
one-year-old widower with four surviving children. Having inherited
his father’s fortune, he was living in style in a London mansion at 1
Dorset Street, Marylebone. Babbage’s soirées were notable occasions—a
heady mix of high society, artists, and scientists. Gatherings of two hun-
dred or more notables were common. One wag had it that wealth alone
was insufficient to procure an invitation. One of three qualifications was
required: ‘intellect, beauty, or rank’.
Lovelace was fascinated by Babbage’s computing devices. At his in-
vitation, she and her mother inspected the working portion of the
Difference Engine. Babbage and Ada struck up a friendship. They cor-
responded regularly, discussing the Analytic Engine and other scientific
topics. At nineteen, Ada married William King to become Augusta Ada
King, Countess of Lovelace.
Not only did Ada Lovelace translate Menabrea’s paper into English,
she extended it with seven notes, more than doubling its length. The
paper—Sketch of the Analytical Engine—was visionary. Although the An-
alytic Engine was never built, Babbage had specified the instructions
that the envisaged machine would perform. This allowed Babbage and
Lovelace to write programs for the non-existent computer.
In the paper, the authors emphasized the relationship between
an algorithm and its equivalent program. They explained that an
algorithm is an abstract calculation method written down, in Babbage
and Lovelace’s case, as a sequence of mathematical equations. The
paper drew out the equivalence between the algorithm expressed as
equations—or algebraic formulae, as they put in—and the program
encoded on punched cards: 28
The cards are merely a translation of algebraic formulae, or, to express it
better, another form of analytical notation.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

44 Computer Dreams

The sequence of mathematical equations could only be performed by

a human. In contrast, the instructions on the punched cards could be
carried out automatically at high speed by the Analytic Engine.
Lovelace’s addendum to the paper presented a series of numerical
algorithms encoded as programs. Lacking an Analytic Engine, the only
way to test a program was to execute it manually, mimicking the
actions of the computer. In Note G of the paper, Lovelace provides
an execution trace of a program for calculating the first five Bernoulli
numbers. The trace lists the instructions as they would have been exe-
cuted by the computer, together with the results obtained after every
step. Unbeknownst to Lovelace, her trace echoed the Old Babylonian
presentation of algorithms in tandem with worked examples. Modern-
day programmers still use traces to better understand the behaviour
of their programs.
Prophetically, the authors noted the prospect of mistakes, or bugs, in
programs: 28
Granted that the actual mechanism is unerring in its processes, the cards
may give it wrong orders.
Ironically, an error has since been found in one of the program listings
included in the paper: a three where there should have been a four. This
was the first bug in a software release. The first of a great many!
The paper also put forth the idea that the computer might process
other forms of information, not just numbers: 28
It might act upon other things besides numbers, were objects found
whose mutual fundamental relations could be expressed by those of the
abstract science of operations, and which should be also susceptible of
adaptations to the action of the operating notation and mechanism of
the engine.
In other words, while the Engine was designed to perform arithmetic,
the symbols employed could represent other forms of information.
Furthermore, the machine could be programmed, or modified, so as
to process these other forms of data. The authors understood that
the machine could manipulate symbols representing letters, words,
and logical values. Babbage even toyed with writing programs to play
Tic-Tac-Toe and Chess.
Babbage, Menabrea, and Lovelace’s spectacular paper was the first
glimmer of a new science. A science at the nexus of algorithms,
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

The Turing Machine 45

programming, and data: the science of computers. It would be another

100 years before the field was recognized and named as computer
science.
This was to be Lovelace’s only scientific publication. She suffered from
poor health for many years, and aged only thirty-six, she died from
cancer. She and Babbage remained close friends until the end. At her
request, she was buried alongside her estranged father in Hucknall,
Nottinghamshire.
Babbage worked intermittently on the Analytic Engine for the rest of
his days. Time and again, his plans were frustrated by the deficiencies of
the mechanical fabrication technologies of the day. Babbage’s Analytic
Engine was another glorious failure.
Nonetheless, Babbage was tireless. He travelled widely. On one oc-
casion, he descended into Mount Vesuvius with the goal of surveying
its volcano. He authored scholarly works on economics, geology, life
assurance, and philosophy. Even aside from his computing engines, he
was a prolific inventor. He dabbled in politics to the extent of stand-
ing for election. In his later years, Babbage became embittered by his
seeming lack of recognition by the establishment. He engaged in a
public campaign of words against the street musicians of London whose
cacophony he found insufferable. Babbage died in 1871, eighteen years
after Lovelace, and his dream of a mechanical digital computer died
with him.
With hindsight, it is clear that the Analytic Engine had all of the
ingredients of the modern computer, except one. It was mechanical,
not electronic. The commercially manufactured electric light bulb was
invented by Thomas Edison after Babbage’s death. It would be another
fifty years before anyone attempted to build a programmable electronic
computer. Babbage’s legacy to computing was Lovelace’s paper.
Without a programmable Engine, it was up to the theoreticians to
chart the future of algorithms and computation. One of the most
influential of these was Alan Turing (Figure 3.2).

The Turing Machine

Turing was born in 1912 in London, England. As a consequence of his
father being a colonial civil servant, Turing’s parents returned to India
when Alan was just one year old. Turing and his brother remained
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

46 Computer Dreams

Figure 3.2 Computing pioneer Alan Turing, c. 1930s.

in England in the care of a retired army colonel and his wife. It was
several years before their mother rejoined her children in England.
The period of familial co-habitation proved short. Turing was sent to
boarding school at thirteen.
At school, Turing befriended classmate Christopher Morcom. The
two shared a deep interest in science and mathematics. They passed
notes discussing puzzles and proofs back and forth in class. Turing
came to worship the ground upon which Morcom walked. Tragically,
Morcom died from tuberculosis in 1930. Turing was deeply affected by
the loss.
In her memoirs, Turing’s mother, Ethel Sara Turing (née Stoney),
recalled her adult son with fondness: 32
He could be abstracted and dreamy, absorbed in his own thoughts, which
on occasion made him seem unsociable [ …]. There were times when his
shyness led him into extreme gaucherie.
Some did not share his mother’s sympathy and saw him as a loner. One
of his lecturers was to speculate that Turing’s isolation and his insistence
on working things out from first principles bestowed a rare freshness on
his work. Perhaps because of his brilliance, Turing did not tolerate fools
lightly. He was also prone to eccentricity, practicing his lectures in front
of his teddy bear, Porgy, and chaining his mug to a radiator to prevent
theft. Turing was a rare combination of difficult to get on with but well-
liked by many of his peers.
Turing won a scholarship to study at the University of Cambridge
and graduated with a first-class honour’s degree in Mathematics. In
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

The Turing Machine 47

Figure 3.3 Turing machine.

the course of further studies at the university, Turing originated three

important ideas in a remarkable scientific paper: he formally defined an
algorithm; he defined the capabilities that a general-purpose computer
must possess; and he employed these definitions to prove that some
functions are not computable. Amazingly, Turing did all of this before
a single digital computer was ever built.
Turing proposed an imaginary computer, now known as a Turing
machine, which consists of an infinite paper tape, a tape head, a mem-
ory, and a set of instructions governing the operation of the machine.
The tape is divided into cells. Each cell has enough space for just one
symbol to be written. The tape head can read or write one symbol at
a time to the cell directly beneath it. The machine can move the tape
back and forth, cell by cell. It can also store one value in memory. The
value stored is referred to as the current state of the machine.
Associated with the machine is a set of instructions, which control
the action of the machine. The way the instructions function is quite
different from how instructions work in a modern computer. In a
Turing machine, every instruction is made up of a pre-condition and
three actions, and the actions are performed if the pre-condition is met.
The pre-condition depends on the current state of the machine (the
value in memory) and the symbol currently underneath the tape head.
If the state and symbol match the values specified in a pre-condition,
then the actions associated with the pre-condition are performed. The
permissible actions are as follows.
1. The tape head can replace, erase, or leave the symbol directly
beneath it unchanged.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

48 Computer Dreams

2. The tape can be moved left or right by a single cell, or remain still.
3. The state in memory can be updated or left unchanged
Turing envisaged that a human programmer would write programs
for the machine to execute. The programmer would provide the pro-
gram and input data to an associate who would manually operate
the machine. With hindsight, it is easy to see that processing the in-
structions is so straightforward that the human operator could be
replaced by a mechanical or electronic device. The operator performs
the following algorithm:

Write the input values as symbols on the paper tape.

Set the initial state of the machine.
Repeat the following steps:
Check the symbol currently under the tape head.
Check the current state of the machine.
Search the instructions to find a matching pre-condition.
Perform the three actions associated with the matching
pre-condition.
Stop repeating when the memory holds the designated halt
state.
When the machine halts, the results are found on the paper tape.
The machine seems antiquated to modern eyes, but all of the
necessary features of a computer are there—reading and writing
data, executing easily modified instructions, processing symbols
representing information, making decisions based on data, and
repeating instructions. Turing never intended that his machine be built.
Rather, it was always meant to be an abstract model of a computing
machine—a conceit that would enable development of the theory of
computation.
Crucially, the Turing machine manipulates symbols. It is up to
humans to ascribe meaning to the symbols. The symbols can be
interpreted as representing numbers, letters, logical (true/false) values,
colours, or any one of a myriad of other quantities.
The Turing machine does not possess dedicated instructions for
arithmetic (i.e. additions, subtractions, multiplications, and divisions).
Arithmetic operations are implemented by executing programs that
process the symbols on the tape so as to achieve the effect of arithmetic.
For example, the calculation 2 plus 2 is performed by a program that
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

The Turing Machine 49

replaces the symbols ‘2+2’ on the tape with the symbol ‘4’. In a modern
computer, arithmetic operations are built-in so as to increase processing
speed.
Turing proposed that his machine was flexible enough to perform
any algorithm. His proposal, which is now generally accepted, is a two-
sided coin. It defines what an algorithm, and a general-purpose com-
puter, are. An algorithm is a sequence of steps that a Turing machine
can be programmed to perform. A general-purpose computer is any
machine that can execute programs equivalent to those which can be
performed by a Turing machine.
Nowadays, the mark of a general-purpose computer is that it is Turing
complete. In other words, it can mimic the operation of a Turing machine.
The paper tape symbols can, of course, be substituted by other physical
quantities, e.g. the electronic voltage levels used in modern computers.
All modern computers are Turing complete. If they weren’t Turing
complete, they wouldn’t be general-purpose computers.
An essential feature of the Turing machine is its ability to inspect
data and to make decisions about what action to perform next. It is this
capability that raises the computer above the automatic calculator. A
calculator cannot make decisions. It can process data, but not respond
to it. Decision-making capability gives computers the power to perform
algorithms.
Turing used his hypothetical machine to assist him in tackling a
classic problem in computability: ‘Can all functions be calculated using
algorithms?’ A function takes an input and produces a value as output.
Multiplication is a computable function, meaning that there is a known
algorithm for calculating the result of a multiplication for all possible
input values. The question that Turing was wrestling with was: ‘Are all
functions computable?’.
He proved that the answer is ‘no’. There are some functions that are
not computable by means of an algorithm. He demonstrated that one
particular function is not computable.
The halting problem queries if there is an algorithm that can always
determine if another algorithm will terminate? The halting problem ex-
presses a practical difficulty in programming. By mistake, programmers
can easily write a program in which some of the steps repeat forever.
This circumstance is called an infinite loop. Typically, it is not desirable
that a program never terminates. It would be helpful for programmers
to have a checker program that would analyse a newly written program
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

50 Computer Dreams

and determine if it contains any infinite loops. In this way, execution of

infinite loops could be avoided.
Turing proved that no general checker algorithm can exist. His proof
hinges on a paradox. For which, the Liar’s paradox is a good example. It
is encapsulated by the declaration:
This sentence is false.
This sentence, like any logical statement, can be either true or false. If
the sentence is true, then we have to conclude that: ‘This sentence is
false’. A statement cannot be both true and false at the same time. It
is self-contradictory. Alternatively, if the sentence is false then we have
to conclude that: ‘This sentence is not false’—another contradiction.
Since both possibilities (true-and-false as well as false-and-true) are
contradictions, the statement is a paradox.
Turing used a paradox to prove that a solution to the halting problem
does not exist. Paradox-based proofs work as follows:

Take a statement that you wish to prove false.

For the moment, assume that the statement is true.
Develop a logical line of reasoning based the assumption.
If the conclusion is a paradox and the logical line of reasoning is
correct,
then the assumption must be invalid.

Turing started with the assumption that:

A checker algorithm that solves the halting problem does exist.
This hypothetical checker algorithm indicates whether a program halts
or not. If the program under test halts then the checker outputs,
‘Halts’. Otherwise, the Checker outputs, “Does not halt”. The checker
algorithm is then:

Take a program as an input.

If the program always halts,
then output “Halts”,
else output “Does not halt”.

Next, Turing followed a logical line of reasoning. He started by

creating a program that runs the checker on itself (Figure 3.4):
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

The Turing Machine 51

Figure 3.4 Turing’s answer to the halting problem.

Repeat the following step:

Run the checker taking the checker as input.
Stop repeating if the checker output is “Does not halt”.

The program examines the output of the checker. If the checker

outputs “Does not halt”, the loop terminates and the program halts. If
the checker outputs “Halts”, the program repeats indefinitely and does
not halt. However, the checker is checking itself. Thus, the checker only
halts if the checker does not halt. Conversely, the checker does not halt
if the checker does halt. Both outcomes are contradictions. There is a
paradox.
Since the logical line of reasoning is correct, the original assumption
must be invalid. This means that a checker algorithm that solves the
halting problem cannot exist. The halting problem is not computable.
There exist functions that, even with complete knowledge, are not
computable. There is a limit to computation.
The good news is that many useful functions—mappings from input
values to output values—are computable. The difficulty is coming up
with efficient algorithms that will perform the desired mapping.
In 1936, Turing moved to Princeton University to undertake a PhD
degree. Whereas a Bachelor of Arts, or Science, degree involves a set of
taught courses and a series of written examinations, the more advanced
Doctor of Philosophy degree consists of a monolithic research project.
A PhD culminates in submission of a doctoral thesis and an intense oral,
or viva voce, examination. The award of a PhD hinges on the candidate
producing a novel piece of work supported by a proof, be it experi-
mental or mathematical. Turing’s PhD, supervised by the American
mathematician and logician Alonzo Church, focused on the theoretical
problems of computation.
Turing returned to Cambridge University in 1938. On 4 September
1939, the day after the British Declaration of War on Nazi Germany,
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

52 Computer Dreams

Turing joined the Government Codes and Ciphers School at Bletchley

Park. Codenamed Station X, Bletchley Park was the UK’s top-secret
wartime code-breaking centre.
Based on intelligence received from a Polish group, Turing and Gor-
don Welchman developed a special-purpose computer to assist in de-
crypting the German Enigma codes. The Bombe—an electromechan-
ical computing device—was completed in 1940. Although it could be
reconfigured, the Bombe was not programmable. It, like other devices
of its generation, could only perform a fixed algorithm. The Bombe
allowed the team to decode intercepted radio communications from
German submarines. The intelligence gleaned enabled the British Navy
to determine German U-boat positions and forewarn allied shipping
of imminent attack. As a direct consequence, the lives of many allied
seamen were saved. Turing was awarded the Order of the British Empire
in recognition of his wartime service. Details of his top-secret activities
were not disclosed.
After the war, Station X was disbanded. Turing joined the National
Physical Laboratory (NPL) in London, where he embarked on a project
to design a general-purpose electronic computer. Due to the difficulty
of the task and Turing’s limitations in working with others, progress
was slow. Frustrated, Turing left the NPL to return to Cambridge before
taking a job at Manchester University. Manchester had pressed ahead
with development of its own electronic computer.
Having spent his career contemplating the prospect of a computer,
Turing finally had his hands on one. He took to programming the
Manchester University Computer and wrote a manual for it. He au-
thored a series of scientific papers pondering possible future applications
of computers. Foreshadowing modern bioinformatics, he suggested
that computers could predict how molecules might behave in biolog-
ical systems. He suggested that computers could tackle problems that
had hitherto required human intelligence. He speculated that by the
year 2000, it would be impossible for a human interrogator, communi-
cating by text message, to distinguish between a human subject and a
computer. This so-called Turing test for artificial intelligence has yet to
be passed.
In 1952, Turing reported a burglary at his home. He informed the
police that a friend of his—Arnold Murray—knew the burglar. When
pressed, Turing acknowledged a homosexual relationship with Murray.
Criminal proceedings were brought against Turing for the crime of
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

The Turing Machine 53

‘gross indecency’. The court sentenced Turing to hormone ‘therapy’—a

euphemism for chemical castration. Previously a keen amateur athlete
and marathon runner, the ‘therapy’ seems to have had a negative effect
on his health.
Two years later, at the age of just forty-one, Turing was discovered
dead in his own bed. Toxicology tests indicated that the cause of death
was cyanide poisoning. The coroner’s inquest concluded that Turing
died by suicide. Many have questioned that verdict. There was no note.
A half-eaten apple was found beside Turing’s bed, but it was never tested
for cyanide. His friends testified that he was in good spirits in the days
prior to his death. Some have suggested that Turing was assassinated
by agents of the state seeking to prevent him from revealing wartime
secrets. The speculation seems far-fetched. Turing habitually conducted
chemical experiments at home. It may well be that his death was simply
an accident.
The goings-on at Station X were hushed up until the 1970s. Even
the families of the men and women who worked there knew nothing
of their achievements at Bletchley Park. Belatedly, in 2013, Turing was
granted a posthumous pardon for his ‘crime’ of homosexual activity by
Queen Elizabeth.
Turing’s legacy to computer science is immense. He defined comput-
ers and algorithms; he set limits on computation; his Turing machine
remains the benchmark against which all computers are compared;
passing the Turing test has become one of the all-time goals of artificial
intelligence. More important than his inventions, though, are the small
nuggets of ideas that he scattered throughout his papers. His seemingly
offhand musings opened up whole avenues of future enquiry for those
following in his wake. In recognition of his achievements, the greatest
honour in computer science is the ACM Turing Award. The annual
one million dollar prize is offered by the Association of Computing
Machines (ACM) for contributions ‘of lasting and major technical im-
portance’.
While Turing was busy cracking the Enigma codes at Bletchley Park,
programmable and (mostly) electronic computers were under devel-
opment elsewhere. In Berlin, the German engineer Konrad Zuse built
a series of relay-based electromechanical computing devices. The pro-
grammable and fully automatic Z3 was completed in 1941. Although
not Turing complete, the Z3 contained many advanced features,
which would later be re-invented elsewhere. Zuse’s efforts were greatly
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

54 Computer Dreams

hindered by the war. Lack of parts, limited funds, and aerial bombing
all played their part in slowing development of the Z series. By 1945,
Zuse’s work was practically at a standstill. Development of the Turing
complete Z4 computer was fatefully stalled. It would be 1949 before
Zuse re-established a company to manufacture his devices. The Z4 was
finally delivered to ETH Zurich in 1950. Zuse’s company built over 200
computers before it was acquired by German electronics conglomerate
Siemens.
In the US, the Second World War was a boon to the development of
the first computers. Inspired by a demonstration model of Babbage’s
Difference Engine, Howard Aiken designed an electronic computer at
Harvard University. Funded and built by IBM, the Harvard Mark I (AKA
the Automatic Sequence Controlled Calculator) was delivered in 1944.
Fully programmable and automatic, the machine could run for days
without intervention. However, it lacked decision-making capability,
and as such, the Harvard Mark I was not Turing Complete and, hence,
not a general-purpose computer.
The world’s first operational digital general-purpose computer was
constructed in Pennsylvania. Bankrolled by the US Army, and with
a clear military application, the machine was unveiled to the assem-
bled press in 1946. It was the beginning of a revolution in algorithm
development.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

4
Weather Forecasts

For six days and seven nights

The wind blew, flood and tempest overwhelmed the land;
When the seventh day arrived the tempest, flood and onslaught,
Which had struggled like a woman in labour, blew themselves out.
The sea became calm, the imhullu-wind grew quiet, the flood held back.
Unknown author, translated by Stephanie Dalley
The Epic of Gilgamesh, circa 2,000 bce 2

Since time immemorial, lives have depended on the vagaries of the

weather. Frequently, human catastrophe could have been averted, if
only the weather were known a day in advance. In 2,000 bce, foretelling
the weather accurately was the preserve of the gods. Seafarers and
farmers looked to weather lore and omens for guidance.
About 650 bce, the Babylonians attempted to forecast the weather
more precisely based on observation of cloud formations. Around
340 bce, the Greek philosopher, Aristotle, wrote Meteorologica, the first
significant book on the nature of weather. His immediate successor
at the Lyceum compiled an accompanying text, Theophrastus of Eresus
on Winds and On Weather Signs, which documented weather lore. These
two books remained the definitive statements on the subject for
almost two thousand years. Unfortunately, both were fundamentally
wrong.
In the post-Enlightenment years, scientists took to painstakingly
recording measurements of weather conditions. Several countries
established centralized metrological services. The UK founded the
Meteorological Office in 1854 and, six years later, the US Weather
Bureau commenced operations. Facilitated by the recent invention
of the electrical telegraph, these services collected meteorological data
from outlying regions and, based on this information, issued weather
forecasts. Their forecasting method was basic. Meteorologists simply
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

56 Weather Forecasts

searched the historical record for the closest approximation to current

conditions. They then predicted that the weather would develop in
the same manner once again. Sometimes, this approach worked. On
occasion, it was disastrously incorrect.
The first inkling of a more accurate forecasting method came at the
beginning of the twentieth century. Cleveland Abbe, Head of the US
Weather Bureau, pointed out that the Earth’s atmosphere is essentially
a mixture of gases. He contended that gases must behave in the same
way in the atmosphere, as in the laboratory. Contemporary scientists
could predict what gases in the laboratory would do when subjected to
heat, pressure, and movement. Surely meteorologists could use these
same scientific laws to predict what gases in the atmosphere would do
when subjected to the heat of the sun and the flow of the wind? The laws
of hydrodynamics (forces acting on fluids) and thermodynamics (heat
acting on fluids) were known. Why not apply them to the atmosphere?
Shortly thereafter, Norwegian scientist Vilhelm Bjerknes suggested
a two-step method for forecasting. In the first step, the current state
of the weather is measured. In the second, a set of equations are used
to predict the future pressure, temperature, density, humidity, and
wind velocity in the atmosphere. Bjerknes derived these equations from
the known laws of physics. Due to their complexity, he was unable
to wrangle a straightforward solution from the equations. Instead, he
resorted to charts and graphs to convert observations to estimates of
future conditions. Each chart was designed to advance the weather by
a fixed time unit. By repeating these steps, Bjerknes argued, weather
predictions could be projected into the future. The conditions output
from the first iteration could be the input to the second and so on.
While Bjerknes’ iterative method was a breakthrough, the accu-
racy of his approach was limited by the charts and graphs. A reliable
method for calculating future conditions from current measurements
was needed. The problem was fraught with difficulty. The equations
were complex and seemingly impossible to solve.
A man with little previous experience of meteorology decided to take
up the challenge.

Numerical Forecasts
Born in Newcastle, England in 1881, Lewis Fry Richardson (Figure 4.1)
studied science at Newcastle University and King’s College, Cambridge.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Numerical Forecasts 57

Figure 4.1 Numerical weather forecaster Lewis Fry Richardson, 1931.

(© National Portrait Gallery, London.)

After graduation, he held a series of short-term research positions

working on fluid flow and differential problems. In 1913, he was ap-
pointed superintendent of the remote Eskdalemuir weather observa-
tory. Located in the uplands of southern Scotland, the Eskdalemuir
landscape is beautiful, wind-swept, and stark. Richardson’s duties com-
prised of recording the weather, monitoring seismic vibrations, and
noting variations in the Earth’s magnetic field. Thankfully, the job came
with a house and plenty of spare time. Amidst ‘the bleak and humid
solitude of Eskdalemuir’, Richardson embarked on a mission to develop
and test a numerical algorithm for weather forecasting. Richardson’s
experiment was underpinned by the inviolable laws of physics.
Richardson divided the atmosphere into an imaginary three-
dimensional grid of cells. A cell might be 100 miles wide by 100 miles
long and two miles high. He assumed that within a cell, the atmosphere
is relatively homogeneous. In other words, all points in the cell
possess approximately the same wind speed, wind direction, humidity,
pressure, temperature, and density. To capture the state of the
atmosphere, he wrote down the values of these quantities for every cell.
Thus, the prevailing conditions were represented as a list of numbers.
To determine how conditions evolved over time, Richardson di-
vided every day into a series of time steps. A time step might be one
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

58 Weather Forecasts

hour. Beginning with the observed conditions, he calculated the likely

state of the weather in the next time step. To make this determina-
tion, he adapted the laboratory-derived gas equations so that he could
calculate the state of a cell based on its own and its neighbours’ state
in the previous time step. Richardson performed these calculations for
every cell. The completed table of values for one time step was then
used as the input for calculation of the conditions in the next time step.
Special equations were used for boundary cells at the top and bottom of
the atmosphere. The heating effects of the sun were incorporated based
on the time of day. Even the effects of the rotation of the Earth were
included. Cell by cell, time step after time step, Richardson calculated
the weather. His algorithm can be summarized as follows:

Measure the initial conditions in very cell.

Repeat the following for every time step:
Repeat the following for every cell:
Calculate the cell’s state from its and its neighbours’
state in the previous time step.
Stop repeating when all cells have been processed.
Stop repeating when all time steps have been processed.
Output the completed forecast.

Today, Richardson’s algorithm would be called a simulation. It predicts,

by means of calculation, how a real-world physical system changes
over time. The equations governing the simulation are a model of the
dynamics of the real-world conditions.
Richardson tested his algorithm using historical weather data mea-
sured over Germany on 20 May 1910. He sought to predict the pressure
and wind speed at two points on the map based on measurements
taken six hours earlier. Richardson rigorously applied his algorithm,
performing every calculation manually. It took months.
In the end, Richardson’s predictions were horribly inaccurate. His
algorithm estimated that the surface pressure would rise to 145 hPa—
a completely unrealistic value. In fact, pressure hardly changed at all
that day. Richardson blamed the discrepancy on errors in how he
represented the initial winds.
Undaunted, Richardson published his findings in 1922 in Weather
Prediction by Numerical Process. In the book, he envisaged a great hall in
which 64,000 human computers aided by mechanical calculators would
compute the weather forecast in real-time.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

ENIAC 59

Richardson’s book was not well received. His algorithm was wildly
inaccurate and outlandishly impractical. It necessitated a vast amount
of computation. The only way forward was with the assistance of a
high-speed calculating machine. It would be almost thirty years before
numerical weather forecasting was revisited.

ENIAC
The first operational general-purpose computer was designed and built
at the University of Pennsylvania during the Second World War. The
ENIAC (Electronic Numerical Integrator And Computer) was designed
by John Mauchly and Presper Eckert, two professors at the college. In a
twist of fate, most of the credit went to world-famous mathematician,
John von Neumann.
Mauchly was born in Cincinnati in 1907. Such was his ability, he was
allowed to begin his PhD studies in Physics before completing the more
basic BSc degree. On graduation, Mauchly was appointed as a Lecturer
at Ursinus College in Pennsylvania.
In 1941, Mauchly took a course on Electronic Engineering at the
Moore School in the University of Pennsylvania. Sponsored by the US
Navy, the course focused on electronics for the military. Eckert, a recent
graduate of the Moore School, was one of the instructors on the course.
Although Eckert hadn’t been an ace student, he was a superb practical
engineer. Despite Mauchly being Eckert’s senior by twelve years, the
two hit it off, bonding over a shared fascination with gadgets. After the
course, Mauchly was hired by the Moore School.
In the shadow of the Second World War, the Moore School was host
to human computers working for the US Army. These human comput-
ers were employed by the Ballistic Research Laboratory (BRL) situated
at the nearby Aberdeen Proving Ground in Maryland. The team, as-
sisted by the Moore School’s mechanical calculators, produced ballistics
tables for the artillery. The tables were used by gunnery officers on
the battlefield to determine the correct firing angles for their artillery
pieces. The tables allowed an officer to take into account the equipment
type, air pressure, wind velocity, wind direction, target range, and target
altitude. The BRL employed one hundred female graduate mathe-
maticians to perform the exacting and time-consuming calculations.
Even with that workforce, the Laboratory was unable to keep up with
demand.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

60 Weather Forecasts

In April 1942, Mauchly wrote a proposal outlining the design of

an electronic computer. His report was circulated within the Moore
School. Lieutenant Herman Goldstine, the leader of the trajectory tabu-
lation effort, heard about the document. A Professor of Mathematics at
the University of Michigan, Goldstine immediately saw the potential for
Mauchly’s computer to alleviate the bottleneck in the ballistics calcu-
lations. He contacted Mauchly. Pleased with what he heard, Goldstine
applied to his Army superiors for funding to enable Mauchly and Eckert
to build the machine.
Work started on the ENIAC in 1943. The youthful and energetic Eck-
ert was appointed Chief Engineer. The mature Mauchly acted as Con-
sultant to Eckert. Goldstine was Project Manager cum Mathematician.
Although intended for computation of artillery tables, ENIAC was a
fully-fledged general-purpose computer. It was programmable, albeit
with wires and plugs. It could perform calculations, store values, and
make decisions. Like Babbage’s Analytic Engine, ENIAC processed dec-
imal numbers. Mostly electronic, with some electromechanical com-
ponents, the machine was immense, weighing roughly twenty-seven
tonnes and covering over 1,500 square feet of floor space. Racks of
electronic circuits were set into floor-to-ceiling cabinets lining the walls
of the Moore School basement. Row upon row of light bulbs and plug
sockets adorned the front of the cabinets. More cabinets, this time on
wheels, shuttled between the banks of equipment. Multitudes of cables
snaked between sockets in seemingly incomprehensible patterns.
Amidst the electronic equipment, a small cadre of programmers
toiled to cajole the temperamental machine into action. The ENIAC
programmers (Figure 4.2) where taken from the ranks of the math-
ematicians working for the BRL. Kathleen McNulty (later Mauchly
and Antonelli), (Betty) Jean Jennings (later Bartik), (Frances) Betty
Holberton (later Snyder), Marlyn Wescoff (later Meltzer), Frances Bilas
(later Spence), and Ruth Lichterman (later Teitelbaum) figured out
how to program the bewildering machine. It was a nightmare. Multiple
units operated simultaneously. The outputs of each had to be re-timed
so as to synchronize the data transfer. Components kept breaking
down. To make the task somewhat easier, Adele Goldstine (née Katz),
Goldstine’s wife and an instructor at the School, compiled the first
operating manual for the machine.
As ENIAC neared completion, Mauchly and Eckert began to con-
template its successor. In August 1944, they proposed an improved
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

ENIAC 61

Figure 4.2 The ENIAC team, 1946. Left to right: Homer Spence; J. Presper
Eckert, chief engineer; Dr John W. Mauchly, consulting engineer; Elizabeth
Jennings (aka Betty Jean Jennings Bartik); Capt. Herman H. Goldstine, liaison
officer; Ruth Lichterman.

device—the EDVAC (Electronic Discrete Variable Automatic Com-

puter). Again, BRL approved funding. Work on EDVAC commenced
shortly thereafter. At roughly the same time, Goldstine introduced a
new collaborator to the project.
John von Neumann was born in Budapest, Hungary, in 1903. From
a well-to-do family, he was privately educated until the age of ten. At
secondary school, he showed a special aptitude for mathematics, to the
extent that he wrote his first research paper before he was eighteen. Von
Neumann went on to study mathematics at the University of Budapest.
At the same time, he completed a Chemistry degree in Zurich, Switzer-
land. He didn’t bother attending the course in Zurich. He just turned
up for the exams. He would come to be fluent in English, German, and
French, with passable Latin and Greek.
Von Neumann was appointed Professor at Princeton University in
1931. There, he joined Einstein as a member of the Institute for Ad-
vanced Study (IAS). Leon Harmon, who worked at the IAS, described
von Neumann as: 45
a true genius, the only one I’ve ever known. I’ve met Einstein and Oppen-
heimer and Teller and a whole bunch of those other guys. Von Neumann
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

62 Weather Forecasts

was the only genius I ever met. The others were super-smart and great
prima donnas. But von Neumann’s mind was all-encompassing. He could
solve problems in any domain and his mind was always working, always
restless.
With a warm and friendly personality, von Neumann was well liked.
Everyone knew him as ‘Johnny’. He had the humility to listen and the
ability to absorb ideas. Given to sharp suits and fast cars, Johnny was
possessed of an earthy sense of humour. The great intellectual delighted
in people and gossip.
During the Second World War, von Neumann was granted a leave
of absence from Princeton to contribute to military projects. He was
heavily involved in the Manhattan Project, assisting in the design of
the first atomic bomb. The Project demanded a great many calcula-
tions. Von Neumann saw the need for a machine that could calculate
faster than any human. In 1944, von Neumann met Goldstine, by
chance it seems, on a train station platform in Aberdeen, Maryland.
Goldstine introduced himself. The two got to talking and, perhaps to
impress von Neumann, Goldstine mentioned his work on ENIAC. Von
Neumann’s interest was piqued. Goldstine extended an invitation and,
subsequently, von Neumann joined the ENIAC project as a consultant.
Eckert later said: 47
Von Neumann grasped what we were doing quite quickly.
In June 1945, von Neumann wrote a 101-page report entitled First Draft
of a Report on the EDVAC. The report described the new EDVAC design
in detail but neglected to mention Mauchly and Eckert, the machine’s
inventors. With Goldstine’s approval, the report was distributed to
people associated with the project. Von Neumann, sole author of the
first report on EDVAC, was widely seen as the originator of the design.
Eckert complained: 47
I didn’t know he was going to go out and more or less claim it as his own.
He not only did that, but he did it at the time when the material was
classified, and I was not allowed to go out and make speeches about it.
Completed in 1945, ENIAC arrived too late to assist in the war effort.
The giant machine was unveiled to the public on St Valentine’s Day 1946
at a press conference in the Moore School. One of the team, Arthur
Burks, gave a demonstration of ENIAC’s capabilities. He started the
show by adding 5,000 numbers together in one second. Next, Burks
explained that an artillery shell takes thirty seconds to travel from gun
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

ENIAC 63

to target. In contrast, manual calculation of its trajectory takes three

days. He informed the audience that the ENIAC would now perform
the same computation. With a touch of theatre, the main lights were
switched off so that the assembled reporters could more clearly observe
the machine’s flickering lights. Twenty seconds later—faster than the
shell could fly—the calculation was complete.
That evening, the project luminaries gathered for a celebratory din-
ner. The dinner was for the top brass and electronic engineers only.
ENIAC’s female programmers were not invited. 50 It would be fifty years
before the ENIAC programmers received a modicum of the recognition
that they so richly deserved.
The morning headlines were euphoric: 51
ARMY’S NEW WONDER BRAIN AND ITS INVENTORS
ELECTRONIC ‘BRAIN’ COMPUTES 100-YEAR PROBLEM
IN 2 HOURS
ENIAC was moved to its owner’s premises, the BRL facility in Ab-
erdeen, in 1947. The machine remained in use until 1955. Disgruntled,
Mauchly and Eckert, resigned from the Moore School in 1947 to found
their own computer company. The nascent corporation soon ran into
financial distress and was acquired by Remington Rand in 1950.
Controversially, Eckert and Mauchly’s patent application for the
ENIAC was disallowed. The judge’s decision rested, in part, on the
prior existence, and Mauchly’s knowledge of, the Atanasoff-Berry Com-
puter (ABC). Developed by John Atanasoff, a professor at Iowa State
University, and his student, Clifford Berry, the ABC was electronic.
However, the ABC was not programmable and lacked decision-making
capability. By modern standards, the ABC was certainly not a general-
purpose computer. It was a special-purpose electronic calculator. With
hindsight, Judge Larson’s ruling is perplexing. ENIAC was far more
advanced than the ABC and contained many innovative features.
Von Neumann’s paper stripped Mauchly and Eckert of fame. The fail-
ure of their start-up and patent application deprived them of fortune.
After a career in computing, Mauchly passed away in 1980. Eckert, the
younger of the pair, worked at Remington Rand and its successors, until
he died in 1995.
Although ENIAC was designed for ballistics calculations, it seems
that the first operational runs were for a higher purpose. The first runs
were secret calculations for the Manhattan Project. At von Neumann’s
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

64 Weather Forecasts

suggestion, a group from Los Alamos visited ENIAC in 1945. Impressed,

they lobbied for ENIAC to be made available to aid in the computations
needed for design of a hydrogen bomb. Their requests were granted, and
an ongoing relationship was established between the Manhattan Project
and the ENIAC team. This collaboration allowed testing of one of the
most powerful algorithms in history.

Monte Carlo
Stanislaw Ulam (Figure 4.3) was born to a well-off Polish-Jewish family
in 1909. He studied mathematics, graduating with a PhD from the
Lviv Polytechnic Institute in the Ukraine. In 1935, he met John von
Neumann in Warsaw. Von Neumann invited Ulam to work with him for
a few months at the IAS in Princeton. Soon after joining von Neumann,
Ulam procured a lecturing job at Harvard University in Boston. He
moved permanently to the US in 1939, narrowly avoiding the outbreak
of the Second World War in Europe. Two years later, he became a citizen
of the United States. Ulam forged a reputation as a talented mathe-
matician and, in 1943, was invited to join the Manhattan Project in Los
Alamos, New Mexico. The high-powered, collaborative environment at
Los Alamos suited Ulam. Nicholas Metropolis, a Los Alamos colleague,
later wrote of Ulam: 52
His was an informal nature; he would drop in casually, without the
usual amenities. He preferred to chat, more or less at leisure, rather
than to dissertate. Topics would range over mathematics, physics, world
events, local news, games of chance, quotes from the classics—all treated
somewhat episodically but always with a meaningful point. His was a
mind ready to provide a critical link.
In the push to build the bomb, Ulam was assigned the problem of cal-
culating the distance that neutrons (charge-free particles at the centre
of an atom) travel through a shielding material. The problem seemed
intractable. Neutron penetration depends on the particle’s trajectory
and the arrangement of atoms in the shielding material. Imagine a table
tennis ball carelessly launched at a million skittles placed at random.
How far does the ball travel, on average? There are so many possible
paths, how could anyone answer the question?
While in hospital convalescing from an illness, Ulam took to playing
Canfield Solitaire, a single-player card game. Canfield Solitaire uses the
normal deck of fifty-two playing cards. Cards are dealt one by one and
moved between piles according to the rules of the game and the player’s
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Monte Carlo 65

Figure 4.3 Stanislaw Ulam, inventor of the Monte Carlo method, c. 1945. (By
Los Alamos National laboratory. See Permissions.)

decisions. The goal is to end up with just four piles of cards. Each pile
should contain all of the cards from one suit.
The rules are quite simple. When a card is drawn, there is a small
number of legal moves to choose from. In most cases, selecting the best
move is straight-forward.
Ulam wondered, what were his chances of winning a game? Whether
he won or lost depended on the order in which the cards were dealt.
Some sequences of cards lead to a win, others a loss. One way to
calculate the odds was to list all possible card sequences and count the
percentage that lead to wins.
Since a deck contains fifty-two cards, there are fifty-two possible first
cards (Figure 4.4). After that, there are fifty-one cards in the pack, so
there are fifty-one possible second cards. Thus, the number of possible
first and second card sequences is 52 × 51 = 2,652. Extending this calcu-
lation to the whole pack, gives 52 × 51 × 50 × 49 × · · · × 1. This is equal
to an 8 followed by sixty-seven digits. No one could possibly play that
many games.
Ulam wondered if the problem could be simplified. What if he played
just ten games? He could count the percentage of wins. That would give
him an indication of the true probability of winning. Of course, with just
ten games there is the possibility of a lucky streak. This would distort
the odds. What about one hundred games? A lucky streak that long
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

66 Weather Forecasts

Figure 4.4 Possible card deals in Canfield Solitaire. The circled outcomes are
sampled using the Monte Carlo method.

is less likely. Ulam concluded that he could get a reasonable estimate

of the true probability of winning by averaging the outcomes over a
sufficiently large number of games. The key point was that he did not
have to play all possible games. He just had to play enough to get a
reasonable estimate of the true odds.
Still, even one thousand games would take a long time to play.
Ulam realized that a computer could be programmed to play that
many games. The computer could be programmed to ‘deal’ the cards
randomly and play in the same way as Ulam. With enough games, the
percentage of wins would be a reliable estimate for the true winning
probability.
In summary, Ulam’s algorithm operates as follows:

Set the win count at zero.

Repeat the following:
Take a fresh pack of cards.
Repeat the following:
Draw a card at random.
Play the card in the best way.
Stop repeating when the pack is empty.
If the game was won, then add one to the win count.
Stop repeating after a large number of games.
Output the percentage of wins.

Ulam figured that his algorithm would work for more than just
card games. It would also work for the neutron diffusion problem.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Computer Forecasts 67

The neutron trajectories and the shielding atom positions could be

represented by random numbers. The penetration distance could be
calculated for each trajectory and shielding configuration. Averaging
over a large number of random trials would give an estimate of the
typical real-world neutron penetration distance.
Ulam proposed the idea to von Neumann and suggested that the
neutron penetration calculations be run on ENIAC. According to
colleague Nicholas Metropolis: 55
Von Neumann’s interest in the method was contagious and inspiring. His
seemingly relaxed attitude belied an intense interest and a well-disguised
impatient drive.
ENIAC was quickly tasked with testing Ulam’s new method. The results
were considered ‘quite favourable’—a polite understatement. The card
players back at Los Alamos named the new algorithm the Monte Carlo
method after the famous casino in Monaco.
Metropolis and Ulam published the first paper on the Monte Carlo
method in 1949. The method has gone on to become a staple of com-
puter simulation. It allows scientists to estimate the probable outcomes
of complex physical events by randomly sampling a large number
of cases. Today, the Monte Carlo method is intrinsic to speculative
studies in fields as diverse as physics, biology, chemistry, engineering,
economics, business, and law.

Computer Forecasts
After the War, von Neumann returned to academic life at the Institute
for Advanced Study in Princeton (Figure 4.5). There, he initiated a
project to build a new electronic computer along the lines of the ED-
VAC. The IAS computer, as it became known, was to be von Neumann’s
gift to computing. The machine was operational from 1952 until 1958.
More importantly, von Neumann distributed plans for the IAS machine
to an array of research groups and corporations. The IAS machine
became the blueprint for computers all over the world.
In addition, Von Neumann pondered the tasks that computers might
be put to. Perhaps as a result of his work on fluid flow in the 1930s, von
Neumann seems to have been aware of Richardson’s work on numerical
weather forecasts. Inspired, von Neumann secured a grant from the
US Navy for establishment of the first computer meteorology research
group. To kick start the initiative, von Neumann organized a conference
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

68 Weather Forecasts

Figure 4.5 Mathematician John von Neumann with the IAS computer, 1952.
(Photograph by Alan Richards. Courtesy of the Shelby White and Leon Levy Archives Center at
the Institute for Advanced Study, Princeton, NJ, USA.)

to bring together the leading meteorological researchers. The group

took on the challenge of conducting the first weather forecast per-
formed by a computer.
One of the group, Jule Charney, joined von Neumann at the IAS.
Charney tamed the complexity of the gas equations, rendering them
amenable to computer execution. The IAS computer wasn’t quite
ready, so von Neumann again requested time on ENIAC.
On the first Sunday of March 1950, five meteorologists arrived at
the BRL in Aberdeen to perform the forecast. The meteorology and
programming teams worked around the clock, in eight-hour shifts,
for almost the entirety of the next thirty-three days. Forecasts were
computed for North America and Europe on four days in January and
February 1949. The dates were specially selected due to the presence
of significant weather systems. The forecasts predicted barometric
pressure over twenty-four hours. Data from the US Weather Bureau
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Chaos 69

provided the initial conditions and allowed evaluation of the resulting

forecasts. The model was based on a rectangular grid of fifteen by eigh-
teen cells, each 736 km across, with one-hour time steps. ENIAC per-
formed the one million calculations needed for each forecast in about
twenty-four hours—just keeping pace with the unfolding weather.
The results were mixed. Some features were predicted accurately.
Others, such as the position and shape of a cyclone, were incorrect.
The researchers put the errors down to the large cell sizes and the
limitations of the equations. Nonetheless, the concept of numerical
weather prediction by computer was proven. The meteorologists and
programmers just had to work out the details.
Richardson was vindicated at last. Charney mailed a copy of the final
paper describing the 1950 ENIAC forecasting experiments to Richard-
son. The pioneer of numerical forecasting replied, congratulating the
team. In a self-deprecating aside, Richardson remarked that the out-
come of the ENIAC experiments was an: 40
. . . enormous scientific advance on [his own] single, and quite wrong
result.
Richardson passed away in 1953, just two years later.
John von Neumann died in 1957 at the age of fifty-three after a long
battle with cancer. Hans Bethe, a Nobel Laureate, remarked: 60
I have sometimes wondered whether a brain like von Neumann’s does
not indicate a species superior to that of man.
Von Neumann’s obituary was written by Stanislaw Ulam.
Ulam went on to make significant contributions to nuclear physics,
bioinformatics, and mathematics. He held professorial appointments at
a series of prestigious American universities, while working summers at
Los Alamos. Ulam passed away in Santa Fe, New Mexico in 1984.
As a result of refinements in algorithms, advances in computer per-
formance, and increases in the number of weather monitoring stations,
forecasting accuracy steadily improved during the 1950s and 1960s. All
seemed to be progressing well until Edward Lorenz stumbled upon a
fundamental limitation of Richardson’s approach.

Chaos
Edward Lorenz was born in Connecticut in 1917. He studied mathemat-
ics prior to serving as a meteorologist in the US Army Air Corps. He
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

70 Weather Forecasts

went on to study the subject at Massachusetts Institute of Technology

(MIT) before becoming a professor there. His discovery of a problem
with Richardson’s method happened by chance in 1961.
As part of a research project, Lorenz ran a series of weather simula-
tions on a small computer. The exercise should have been routine. What
transpired was distinctly odd.
Lorenz elected to repeat one of the simulations to examine the results
in greater detail. On returning to the computer an hour later, he
compared the new results with the previous outputs. He was startled
to find that the new forecasts were nothing like the old ones. Naturally,
he suspected that the cause was a fault in the computer—a common
occurrence at the time. Before calling computer maintenance, he de-
cided to check the simulation print-outs time step by time step. At the
beginning, they matched. After a while, the values began to diverge.
The discrepancy grew rapidly as the simulation evolved. The difference
between old and new more or less doubled every four simulated days.
By the end of the second simulated month, the old and new outputs
bore no resemblance to one another whatsoever.
Lorenz concluded that both the computer and the program were
working just fine. The discrepancy arose from a very small difference
in the simulation inputs. On the first run, Lorenz had entered the state
of the atmosphere as six digits. On the second run, he had only typed in
three digits. The difference between a six-digit number and the nearest
three-digit number is small. One would except a small difference at the
input to lead to a small difference at the output. On the contrary,
the calculations had caused the difference to grow, ultimately leading
to a large discrepancy between the outputs. Crucially, Lorenz realized
that he wasn’t seeing an artefact of simulation. The simulation was
accurately modelling a real-world phenomenon.
A new science sprang from Lorenz’s accidental discovery. Chaos theory
has since determined that many real-world physical systems are hyper-
sensitive to their initial conditions. Small changes in the initial state
can lead to major differences in conditions later on. The idea was
encapsulated in the popular epithet, ‘the Butterfly Effect’. If conditions
are right, the flap of a butterfly’s wings in Brazil can be the sole cause of a
tornado in Texas a few days later. The example is extreme. Nevertheless,
the idea has proven its worth. Chaos has since been identified in many
real-world systems, including the orbits of asteroids moving within the
rings of Saturn.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Chaos 71

Chaos theory places a bound on the time horizon of numerical

weather forecasts. Small errors in modelling current conditions can
lead to large forecasting errors at a later stage. This time horizon for
accurate forecasting seemed inviolable until Edward Epstein stepped in.
Epstein was born in the Bronx in 1931. Like Lorenz, Epstein was
introduced to meteorology during military service. After leaving the
US Air Force, Epstein worked as a researcher and lecturer at a series
of American colleges. While a visiting scientist at the University of
Stockholm, he published a paper outlining an algorithm that could
mitigate the Butterfly Effect. Epstein’s idea offered a way to extend the
weather forecasting time horizon.
Numerical weather forecasting, as proposed by Richardson, relies
on a single simulation to predict the weather. The simulation starts
with measurements of current conditions and, cell by cell, time step
after time step, calculates how the weather evolves. Epstein’s insight
was to apply Ulam’s Monte Carlo method to Richardson’s numerical
simulation.
Instead of one simulation, Epstein proposed that many simulations be
run. Each simulation begins with randomly perturbed initial conditions.
These initial conditions are created by applying small random changes,
or perturbations, to the observed atmospheric conditions. Given the
limitations of measurement, forecasters don’t know exactly what the
current weather conditions are in any cell. The perturbations allow
forecasters to try out a number of possible scenarios. At the end of
the simulations, the outputs are averaged to obtain a unified final
prediction. This average takes an ensemble of possibilities into account.
Their average is the middle-of-the-road, all-things-considered, most
likely scenario. Epstein’s proposal can be summarized thus:

Measure the current atmosphere conditions.

Repeat the following steps:
Add small random perturbations to the current conditions.
Perform a numeric forecast starting from these initial
conditions.
Store the results.
Stop repeating when sufficient simulations have been
performed.
Output the average forecast.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

72 Weather Forecasts

FLORENCE

Figure 4.6 Hurricane Florence (2018) path predictions obtained using an en-
semble method.

The downside of Epstein’s algorithm is that it requires a lot of calcu-

lation. Running eight Monte Carlo simulations requires eight times the
computer power of a single forecast. For this reason, Epstein’s ensemble
approach wasn’t put into operational use until the early 1990s. Today,
ensemble forecasts are state-of-the-art (see Figure 4.6). For example, the
European Centre for Medium-Range Weather Forecasts now bases its
predictions on fifty-one unique simulations.
Edward Lorenz was to accumulate a long list of scientific prizes for
his work on chaos theory. Edward Epstein devoted his entire career to
meteorology and climate modelling. He was an early advocate for the
concept of man-made climate change. Both passed away in 2008.

Long-Range Forecasts
Like ENIAC, the computers of the 1950s were expensive, power hungry,
and unreliable behemoths. The invention of the transistor and the inte-
grated circuit, in 1947 and 1958 respectively, allowed the miniaturization
of the computer.
Transistors are electronic switches. Containing no moving parts,
other than electrons, transistors are small, low power, reliable, and
incredibly fast. Groups of transistors can be wired together to create
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Long-Range Forecasts 73

logic circuits. These logic circuits can, in turn, be interconnected so as

to create data processing units.
Integrated circuits allow the manufacture of vast numbers of tiny
transistors, together with their metallic inter-connections, at incredibly
low cost. An integrated circuit is inside every computer chip. These chips are
the physical building blocks of modern computers.
Over the years, electronic engineers have refined transistor designs
and integrated circuit technology. In 1965, Gordon Moore, the co-
founder of Intel, noted that his engineering team had managed to
double the number of transistors fabricated on a single integrated circuit
every eighteen months. Moore saw no reason why this trend could not
continue into the future. His prediction, enshrined as ‘Moore’s Law’,
became a roadmap for the industry. Moore’s Law has proven to be one
of the greatest predictions of the modern age. The Law has held true for
more than half a century.
In line with Moore’s prediction, the performance of computers has
risen exponentially. Even as performance has risen, computer size, cost,
and power consumption have plummeted. Today, a top-end computer
chip contains tens of billions of transistors. Moore’s Law is the driving
force behind the vertiginous rise of the computer.
In 2008, Peter and Owen Lynch (University College Dublin and IBM
Ireland) repeated the original ENIAC forecasts on a mobile phone.
Their program—PHONIAC—ran on an off-the-shelf Nokia 6300 mo-
bile phone. ENIAC took twenty-four hours to perform a single forecast.
In contrast, the Nokia 6300 took less than one second. Whereas ENIAC
weighted twenty-seven tonnes, a Nokia 6300 weighs just ninety-one
grams. This is Moore’s Law in action.
Advances in computer technology have enabled new algorithms to
be developed. Algorithms that were once impossibly time consuming
to execute are now routine. Avenues for algorithmic research that once
seemed purely theoretical are now entirely practical. The emergence of
novel computing devices has enabled fresh applications and created the
need for brand new algorithms. Last, but not least, the success of the
computer industry has led to huge increases in the number of people
engaged in software and algorithm development.
Thus, Moore’s Law has also led to exponential growth in the number
of algorithms.
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

5
Artificial Intelligence Emerges

The machine is not a thinking being, but simply an automaton

which acts according to the laws imposed upon it.
Luigi Federico Menabrea and Ada Lovelace
Sketch of the Analytical Engine, 1843 28

In the 1940s and 1950s, a computer was considered to be, in essence, a fast
calculator. Due its high cost and large size, a computer was a centralized,
shared resource. Mainframe computers churned through huge volumes
of repetitive arithmetic calculations. Human operators were employed
as gatekeepers to these new-fangled contraptions, apportioning valu-
able compute time to competing clients. Mainframes ran substantial
data processing jobs one after another with no user interaction. The
final voluminous printouts were presented in batch by the operators to
their grateful clients.
Amidst the expansion of industrial-scale arithmetic, a handful of
visionaries wondered if computers could do more. These few under-
stood that computers were fundamentally symbol manipulators. The
symbols could represent any sort of information. Furthermore, they
opined that, if the symbols were manipulated correctly, a computer
might even perform tasks which had, until that point, required human
intelligence.

More Than Math

Alan Turing left the NPL in 1947 to return to Cambridge University for a
one-year sabbatical. In departing the NPL, Turing abandoned his brain-
child, the Automatic Computing Engine (ACE). The ACE was meant to
be the UK’s first general-purpose computer. However, the project did
not go well. Construction of the machine was exceedingly challenging.
In addition, Turing was difficult to work with. 67 In the wake of his
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

76 Artificial Intelligence Emerges

Figure 5.1 AI pioneer Christopher Strachey. (© Bodleian Library & Camphill

Village Trust. Courtesy of The National Museum of Computing.)

departure, the team pressed on. A much simplified design, the Pilot
ACE, finally became operational in 1950.
That autumn, the group received an unusual request. Christopher
Strachey, a teacher at Harrow School, enquired if he might have a go at
programming the Pilot ACE. Strachey was undoubtedly a novice to pro-
gramming, although in 1950, everyone was a novice to programming.
Born in 1916, Strachey was scion of a well-to-do, intellectual, En-
glish family. He graduated King’s College, Cambridge with a degree
in Physics. In his third year, he suffered a mental breakdown. Later,
his sister attributed the collapse to Strachey coming to terms with his
homosexuality. 68 During the Second World War, Strachey worked on
radar development. Thereafter, he took up employment as a teacher at
Harrow, one of the most exclusive public schools in England.
Strachey’s request was approved and he spent a day of his Christmas
vacation at the NPL, absorbing all the information that he possibly could
about the new machine. Back at Harrow, Strachey took to writing a
program for the Pilot ACE. With no machine at his disposal, Strachey
wrote the program with pen and paper and tested it by imaging the
computer’s actions. Most beginners start with a simple programming
task. Due to either ambition or naivety, Strachey embarked on writing a
program to play Checkers (Draughts in the UK). This was certainly not
an arithmetic exercise. Playing Checkers mandates logical reasoning
and foresight. In other words, playing Checkers requires intelligence.
That spring, Strachey got wind of a new computer at Manchester
University. The project was initiated by Bletchley Park alumnus Max
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

More Than Math 77

Newman just after the war. More powerful than the ACE Pilot, the
Manchester Baby seemed better suited to Strachey’s work. Strachey got
in touch with Turing, who by then was Deputy Director of the Manch-
ester Computing Machine Laboratory. Acquaintances since King’s Col-
lege days, Strachey managed to wheedle a copy of the programming
manual from Turing. Later that summer, Strachey visited Turing to find
out more.
A few months later, Strachey returned to test a program he had
written at Turing’s behest. Overnight, Strachey went from handwritten
notes to a working thousand-line program. The program solved the
problem that Turing had set and, on completion, played The National
Anthem on the computer’s sounder. This was the first music ever played
by a computer. Even Turing was impressed. It was clear that Strachey
was a born programmer.
Strachey was recruited by the National Research and Development
Corporation (NRDC). The NRDC’s remit was to transfer new tech-
nologies from government agencies to the private sector. The NDRC
didn’t have much for Strachey to do at the time, so he continued
programming, and among other things, he invented a program to
compose love letters.
Strachey’s program took a template love letter as input and selected
the adjectives, verbs, adverbs, and nouns at random from pre-stored
lists. From whence came the ardent epistle: 72
Honey Dear
My sympathetic affection beautifully attracts your affectionate enthusi-
asm. You are my loving adoration: my breathless adoration. My fellow
feeling breathlessly hopes for your dear eagerness. My lovesick adoration
cherishes your avid ardour.
Yours wistfully
M. U. C. [Manchester University Computer]
To the bemusement of his colleagues, Strachey pinned the love letters
on the Laboratory noticeboard. While whimsical in nature, Strachey’s
program was the first glimmer of computer creativity.
Strachey finally completed his Checkers program in 1952, describing
it in a paper entitled Logical or Non-Mathematical Programmes.
Checkers is a two-player board game played on the same eight-by-
eight grid as Chess. Players take opposing sides of the board and are given
twelve checkers (disks) each. One player plays white, the other black. To
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

78 Artificial Intelligence Emerges

Figure 5.2 Checkers boards illustrating a simple play by white (left) and a later
jump which removes a black checker (right).

begin, the checkers are placed on black squares in the three rows nearest
the player (Figure 5.2). Players take turns to move a single checker.
Checkers normally move one square diagonally in a forwards direction.
Checkers can jump over an opponent’s neighbouring checker if the
square beyond is unoccupied. A sequence of such jumps can be per-
formed in a single play. All of the opponent’s ‘jumped’ checkers are
removed from the board. The aim of the game is to eliminate all of an
opponent’s checkers. At the start, checkers can only move forwards.
When a piece reaches the far side of the board, it is ‘crowned’ by placing a
checker on top of it. Crowned pieces can be moved diagonally forwards
or backwards.
Checkers is complex. There is no simple strategy that inevitably
leads to a win. Potential plays must be evaluated by imagining how the
game will evolve. A seemingly innocuous play can have unforeseen
repercussions.
Strachey’s algorithm uses numbers to record the position of the
checkers on the board. On its own turn, the algorithm examines all
possible next plays (ten on average). In board game parlance, a move
consists of two plays, one by each player. A single play (or half-move)
is called a ply. For every possible next play, the algorithm assesses its
opponent’s potential responses. This lookahead procedure is applied up
to three moves deep. The results of the lookahead can be visualized as a
tree (Figure 5.3). Every board position is a node, or branching point, in the
tree. Every possible play from that position gives rise to a branch leading
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

More Than Math 79

Figure 5.3 Visualization of the Checkers lookahead tree. Every node is a board
position. Every branch is a play.

to the next board position. The greater the lookahead, the more layers
in the tree. For the nodes at the end of the lookahead, the algorithm
counts the number of checkers that each player retains on the board.
It selects the play at the root of the tree that leads to the greatest
numerical advantage for the computer at the end of the lookahead.
On the Ferranti Mark I, a commercial version of the Manchester
Mark I, every play took one to two minutes of computer time. Even
at that, Strachey’s program wasn’t a particularly good player. With
hindsight, the program’s lookahead depth was insufficient, the decision-
making logic lacked sophistication, and position evaluation was inac-
curate. Nevertheless, the hegemony of computer arithmetic had been
broken. Here was the first working example of artificial intelligence.
Strachey went on to become the University of Oxford’s first Professor
of Computer Science. Unfortunately, while a well-regarded academic,
much of Strachey’s later work remains unrecognized due to his hesi-
tancy to publish academic papers. After a short illness, he passed away
in 1975, aged 58.
Board games have become the barometers of artificial intelligence
(AI). The reasons why are both technical and very human. Board games
have clearly defined goals and rules that are amenable to computer
programming. At any point in time, there is a limited number of options
available to the computer, which makes the problem tractable. Playing
humans is an easily understood indictor of progress. Additionally peo-
ple love contests. A machine that might beat the world champion will
always provoke public interest. Even AI researchers crave an audience.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

80 Artificial Intelligence Emerges

The Trouble with AI

The term ‘artificial intelligence’ was coined in 1955 by John McCarthy in
a proposal submitted to the Rockefeller Foundation. The document re-
quested funding for a two-month, ten-person summer research project
involving luminaries such as Claude Shannon and Marvin Minsky.
McCarthy, an Assistant Professor of Mathematics at Dartmouth College
in New Hampshire, USA, wrote: 74
For the present purpose the artificial intelligence problem is taken to be
that of making a machine behave in ways that would be called intelligent
if a human were so behaving.
The term caught on, but McCarthy’s definition has proven problematic.
There are many aspects of intelligence, but for a moment, let us
consider one specific example. Before 1940, most people would have
said that playing Checkers requires intelligence. The player must un-
derstand the board position and come up with a series of plays that will
lead to a win. In contrast, most people would concur that performing
an algorithm does not require intelligence. Slavishly carrying out one
well-defined step after another is trivial. Even a machine can do that.
Therein lies the rub. Checkers requires intelligence when the algo-
rithm for playing is unknown. As soon as the algorithm is known,
playing Checkers no longer requires intelligence.
AI—as defined by McCarthy—is always the unsolved problem. Once
the algorithm is known, the problem no longer requires intelligence.
Intelligence, we feel, is reserved for intellectual tasks that computers
cannot perform. In ways, AI is akin to stage magic. Once we know how
the trick is done, it is no longer magic.
As algorithms and computers have improved, the boundary of hu-
man intelligence has repeatedly been redrawn. Before calculators, one
might have said that arithmetic requires intelligence. Before Strachey,
one might have claimed that playing board games needs intelligence.
Looking to the future, one wonders where the ultimate boundary
between algorithms and intelligence lies. Perhaps human intelligence,
in its entirety, will prove to be an algorithm.
McCarthy’s definition of AI has caused much confusion. The
layperson imagines that AI is fully formed, equivalent to human
intelligence. In truth, human intelligence is multifaceted and general—
multi-faceted in that there are many aspects to our intelligence. We
can learn, recall, multitask, invent, apply expertise, imagine, perceive,
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Trouble with AI 81

abstract, and so on. Human intelligence is general in that we can

perform a wide variety of tasks. We can make lunch, we can debate,
we can navigate, we can play sports, we can mend broken machines,
and so forth. To be called ‘artificially intelligent’, a computer need only
perform a single task that was previously thought to require human
intelligence. A computer need only play Checkers to be considered in
possession of AI. The proper term for intelligence akin to our own is
human-level artificial general intelligence (HLAGI). HLAGI is what we see in
science fiction movies. Science fact—AI—is very far from HLAGI.
An important aspect of McCarthy’s definition was that he described
AI in terms of outcome. He didn’t require that machines solve
problems in the same way as humans. So long as the human and
machine produce comparable outcomes then he considers AI to be
on a par with human intelligence. For McCarthy, the mechanism does
not matter.
Over the years, the dichotomy between how computers and how
humans perform tasks that require intelligence has engendered much
philosophical debate. The crux of the matter can be best expressed in
the simple query: ‘Can machines think?’ The answer, of course, depends
on what is meant by the word ‘think’. If thought is a biological process
of the brain, then clearly computers cannot think. To most people,
such a requirement seems overly restrictive. Why should the substrate
delineate thinking from non-thinking? If aliens arrived from another
planet, would we deny that they can think simply because they are
silicon-based lifeforms, rather than carbon-based? I think not.
To most people, the prerequisites for thought are intelligence plus
consciousness. Consciousness, the state of being self-aware, allows sen-
tient beings to ‘hear’ themselves ‘think’. For most people, consciousness
is central to thinking. So far, computers are certainly not conscious.
Furthermore, we have no idea how to make them conscious. Thinking
machines are a long way off. Perhaps they are impossible.
McCarthy intimated that the question ‘Can machines think?’ is irrel-
evant. Alan Turing concurred: 36
The original question, ‘Can machines think?’ I believe to be too mean-
ingless to deserve discussion. Nevertheless, I believe that at the end of the
century the use of words and general educated opinion will have altered
so much that one will be able to speak of machines thinking without
expecting to be contradicted.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

82 Artificial Intelligence Emerges

For Turing, if a machine’s behaviour is indistinguishable from human

intelligence then we will conclude that the machine ‘thinks’. To him,
whether it really ‘thinks’, or not, only matters to the philosopher.
In contrast, whether a machine ‘feels’ or not has major ramifications.
If the machine has consciousness and emotion, then surely, we have
ethical responsibilities towards it. Such questions will gain in import as
technology advances.
McCarthy’s initiative was approved for funding by the Rockefeller
Foundation. The Dartmouth Conference went ahead in the summer of
1955. The gathering heralded the beginning of AI as a field of research.
Disappointingly, the Conference proved to be a meandering series of
getting-to-know-you events. Participants drifted in and out of the Con-
ference, touting their own agendas. Few gleaned new insights. Most
opined that the Conference did not achieve much. 45 In retrospect, one
presentation did offer a signpost to the future. Two American scientists,
Allen Newell and Herbert Simon, unveiled a computer program that
could perform algebra.

Machine Reasoning
Algebra is the branch of mathematics concerned with equations that
include unknown values. The unknowns are signified by letters. Using
the rules of algebra, mathematicians seek to re-arrange and combine
equations so that the values of the unknowns can be determined.
For thousands of years, manipulating equations was the domain of
mathematicians. This was something that abaci, calculators, and early
computer programs could not do. In 1946, John Mauchly, co-inventor
of ENIAC wrote: 51
I might point out that a calculating machine doesn’t know how to do
algebra, but only arithmetic.
By the time of the Dartmouth Conference, Newell and Simon were
working at RAND Corporation. Based in Santa Monica, California,
RAND was, and still is, a not-for-profit research institute. Established
after the Second World War, RAND specializes in planning, policy, and
decision-making research for governmental bodies and corporations.
In the 1950s, RAND’s number one customer was the US Air Force.
RAND was a researchers’ paradise—intellectual freedom, smart col-
leagues, healthy budgets, and no teaching. Effectively, RAND employ-
ees were told: 45
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Machine Reasoning 83

Figure 5.4 Designers of the Logic Theorist, Allen Newell and Herbert Simon.
(Courtesy Carneige Mellon University.)

Here’s a bag of money, go off and spend it in the best interests of the Air
Force.
Simon, the elder of the pair by eleven years, was from Milwaukee.
By the 1950s, he was an established political scientist and economist. He
was a member of faculty at Carnegie Institute of Technology (CIT) in
Pittsburgh and spent summers working at RAND.
Newell grew up in San Francisco, California. He graduated with a
degree in Physics from Stanford University before dropping out of an
advanced degree in Mathematics at Princeton to join RAND.
The pair first dabbled in computers while working on projects with
the goal of enhancing organizational efficiency in air defence centres.
RAND’s computer, JOHNNIAC, was based on the IAS blueprint. John
von Neumann himself was a guest lecturer at RAND. Nonetheless, it
was a talk by Oliver Selfridge of MIT Lincoln Labs that captured Newell’s
imagination. At it, Selfridge described his work on recognizing simple
letters (Xs and Os) in images. Newell later reflected: 45
[The talk] turned my life. I mean that was a point at which I started work-
ing on artificial intelligence. Very clear—it all happened one afternoon.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

84 Artificial Intelligence Emerges

Over the course of the next year, Newell and Simon developed an AI
program named Logic Theorist. Newell relocated from Santa Monica
to Pittsburgh so as to work more closely with Simon in CIT. Since CIT
didn’t have a computer, the duo tested their program by gathering a
team of students in a classroom and asking them to simulate the be-
haviour of the machine. The group ‘walked’ through programs, calling
out instructions and data updates along the way. After verification,
Simon and Newell transferred the program to Cliff Shaw in RAND
Santa Monica via teletype. Shaw entered the program into JOHNNIAC
and sent the results back to Pittsburgh for analysis.
The team declared the Logic Theorist operational on 15 December
1955. When the teaching term resumed, Simon was triumphant. He
announced: 45
Over Christmas, Allen Newell and I invented a thinking machine.
Logic Theorist performs algebra on logic equations. A logic equation re-
lates variables to one another by means of operators. Variables are denoted
by letters and can have either true or false values. The most common
logical operations are: ‘=’ equals, ‘AND’, and ‘OR’. For example, if we
allocate the following meanings to the variables A, B, and W:
A = ‘Today is Saturday’
B = ‘Today is Sunday’
W = ‘Today is the weekend’
we can construct the equation:
W = A X OR B
meaning that ‘Today is the weekend’ is true if ‘Today is Saturday’ is true
OR ‘Today is Sunday’ is true, excluding the case that both are true.
By means of algebra, equations such as this can be manipulated so
as to reveal new relationships between variables. The series of manip-
ulations that lead from an initial set of equations to a conclusion is
called a proof. The idea is that if the initial set of equations is valid and
the rules of manipulation have been applied properly, the conclusion
must also be valid. The starting equations are called the premises and the
final conclusion the deduction. The proof provides formal step-by-step
evidence that the deduction is valid given the premise. For example,
given:
W = A X OR B
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Machine Reasoning 85

We can prove that:

A = W AND NOT B
In other words, ‘Today is Saturday’ must be true if ‘Today is the week-
end’ is true and ‘Today is Sunday’ is false.
Humans produce proofs by intuition and experience. Logic Theorist
takes a brute-force approach to finding a proof. It tries all possible
algebraic manipulations on the input statements. It repeats this process
for the resulting equations and so on. If it finds the conclusion that it
is looking for, the search terminates. The program then backtracks and
outputs the chain of transformations connecting the deduction to the
original premises. This chain is presented to the user as the proof.
Ultimately, Logic Theorist provided step-by-step proofs for thirty-
eight of the fifty-two theorems in the classic textbook, Principia Mathe-
matica. Indeed, one of Logic Theorist’s proofs is more elegant than the
textbook version.
In 1959, Newell, Shaw, and Simon introduced a new program. The
General Problem Solver used a similar approach to Logic Theorist.
However, as the name suggests, the new program tackled a much
wider variety of algebraic puzzles, including geometry. To speed up the
search, the General Problem Solver algorithm does not try all possible
manipulations. It prioritizes equations that are similar to the desired
deduction. This means that less time is spent exploring futile paths. Of
course, there is an attendant risk that an important line of reasoning
is ignored, and the desired conclusion never reached. Rule guided, or
heuristic, search is now commonplace thanks to its speed.
Newell, Simon, and Shaw’s work on reasoning was highly influential.
An entire field of AI (symbolic reasoning) grew out of the concept of pro-
cessing logical statements as lists of symbols. Simon even went so far as
to claim that the General Problem Solver mimicked human reasoning.
Certainly, there are similarities in the manner in which humans some-
times derive formal mathematical proofs by trial-and-error. However,
human reasoning appears more intuitive and less rigorous than the
General Problem Solver approach.
Newell, the Princeton dropout, was finally awarded a PhD at CIT. In
1967, CIT merged with the Mellon Institute to create the Carnegie Mel-
lon University (CMU). Newell and Simon went on to build one of the
world’s leading AI research groups at CMU. In 1975, they were awarded
the ACM Turing Award for their work on AI and cognitive psychology.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

86 Artificial Intelligence Emerges

Three years later, Simon received the Nobel Prize for contributions to
microeconomics, his other research interest. Newell and Simon lived
the rest of their lives in Pittsburgh. Newell passed away in 1992 (aged
sixty-five) and Simon in 2001 (aged eighty-four).

Machine Learning
The ability to learn is central to human intelligence. In contrast, early
computers could only store and retrieve data. Learning is something
entirely different. Learning is the ability to improve behaviour based
on experience. A child learns to walk by copying adults and by trial-
and-error. Unsteady at first, a toddler’s co-ordination and locomotion
gradually improve until the infant becomes a proficient walker.
The first computer program to display the ability of learn was un-
veiled on public television on 24 February 1956. That program was
written by Arthur Samuel, of IBM. Samuel’s program, like Strachey’s,
played Checkers. The TV demo was so impressive that it was credited
with a fifteen-point uptick in IBM’s share price the next day.
Samuel was born in Kansas in 1901. He received a Master’s degree
in Electronic Engineering from MIT prior to taking up employment
with Bell Labs. After the Second World War, he joined the University of
Illinois as a Professor. Even though the university lacked a computer,
Samuel started work on a Checkers-playing algorithm. Three years
later, after joining IBM, Samuel finally got his hands on a real computer.
At much the same time as Strachey published his paper on Checkers,
Samuel got the first versions of his game-playing program working.
On first sight of Strachey’s paper, Samuel felt that his own work had
been scooped. On closer inspection, it was clear that Strachey’s program
was a weak Checkers player. Confident that he could do better, Samuel
pressed on.
In 1959, Samuel finally published a paper describing his new Checkers
program. The understated title—Some Studies in Machine Learning using the
Game of Checkers—belied the importance of his ideas.
Samuel’s algorithm is more thorough in evaluating positions than
Strachey’s. It achieves this by means of a clever scoring algorithm.
Points are given for various on-board features. A feature is anything that
indicates the strength, or weakness, of a position. One feature is the
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Machine Learning 87

Figure 5.5 Originator of machine learning, Arthur Samuel, 1956. (Courtesy of

International Business Machines Corporation, © International Business Machines Corporation.)

difference in the number of checkers that the two players have on the
board. Another is the number of crowners. Yet another is the relative
positions of checkers. Strategic elements, such as freedom to move or
control of the centre of the board, are also considered to be features.
Points are scored for every feature. The points for a given feature are
multiplied by a weight. The resulting values are totalled to give an overall
score for the position.
The weights determine the relative importance of each feature.
Weights can be positive or negative. A positive weight means that the
feature is beneficial for the computer player. A negative weight means
that the feature reduces the computer’s chances of winning. A large
weight means that a feature has a strong influence on the total score.
Multiple features with low weights can, however, combine to influence
the overall score and thus the final decision.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

88 Artificial Intelligence Emerges

In summary, Samuel’s position evaluation method works as follows:

Take a board position as input.

Set the total score to zero.
Repeat the following for each feature:
Measure the feature on the board.
Calculate the points for the feature.
Multiply by the feature weight.
Add the result to the total score.
Stop repeating when all features have been scored.
Output the total score.

This scoring mechanism is crucial to Samuel’s algorithm. The more

accurately the scores reflect the computer’s chances of winning, the
better the decisions the program makes. Selecting the best features for
analysis is important. But beyond that, determining the best weights is
essential. However, finding the best values for the weights is tricky.
Samuel designed a machine learning algorithm to determine the opti-
mum weights. Initially, the algorithm guesses the weights. The com-
puter then plays a large number of games against itself. One copy of the
program plays the white checkers, the other black. As play proceeds, the
algorithm adjusts the weights so that the calculated scores more accu-
rately predict the game outcome. If the program wins, the weights that
contributed positively to the decision are increased slightly. Similarly,
any weights that contributed negatively are decreased. This reinforces the
winning behaviour. In effect, it encourages the program to play more
like this in the future. If the game is lost, the opposite happens. This
discourages the program from playing in the same way in the following
games. Over a great many games, the learning algorithm fine-tunes the
program’s play.
The advantages of Samuel’s algorithm over manual selection of the
weights are twofold. First, the computer never forgets—every sin-
gle play influences the weights. Second, the computer can play far
more games against itself than any human can ever play. Thus, far more
information is available to the learning process.
Samuel’s development of machine learning was game changing.
Previously, altering the behaviour of a program required manual mod-
ification of the list of instructions. In contrast, the decisions made by
Samuel’s program are controlled by the weights, which are simple
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Machine Learning 89

numerical values. Thus, the behaviour of the program can be adjusted

by changing the weights. The program code does not need to be modi-
fied. Whereas altering program code is difficult, changing a few numeric
weights is trivial. It can be done by an algorithm. This ingenious concept
enabled the automation of learning.
In addition, Samuel included a minimax procedure for selecting plays.
The algorithm performs a lookahead procedure to generate a tree of all
possible future moves (Figure 5.6). Scores are calculated for all of the
boards at the end of the lookahead. Some of the highest scoring boards
are unlikely to occur in a real game since they result from particularly
poor play by the opponent. One should assume that both players make
good plays. To allow for this, the algorithm backtracks over the tree
of plays. The program starts at the leaves of the lookahead tree. On
its own turns, the backtracking algorithm selects the play that leads
to the maximum score. On its opponent’s turns, the program selects
the play that leads to the minimum score. At every decision point, the
score associated with the chosen play is carried backwards to the parent
node. When this backtracking process reaches the root of the tree, the
program makes the play associated with the highest backtracked score.
The minimax procedure functions as follows:

Take the tree of possible plays as input.

Start at the penultimate layer.
Repeat the following for every layer:
Repeat the following for every node in the layer:
If it is the computer’s turn,
then select the play giving the maximum score,
else select the play giving the minimum score.
Copy the minimax score to the current node.
Stop repeating when all nodes in this layer have been
examined.
Stop repeating when the root of the tree is reached.
Output the play with the maximum backtracked score.

Image a simple two-ply lookahead (Figure 5.6). The tree includes the
computer’s potential next plays and its opponent’s possible responses.
The scores at the leaves of the tree (ply 2) are inspected to find the
minimum in each sub-tree. This reflects the opponent’s selection of the
best play from their point of view. These minimum scores are copied
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

90 Artificial Intelligence Emerges

Figure 5.6 Lookahead tree showing backtracking scores obtained using the
minimax algorithm.

to the nodes immediately above (ply 1). This puts the scores 1, 3, 7, 5,
and 6 on the nodes in ply 1. Now, the algorithm selects the play giving
the highest score. This means that the computer chooses the best play
from its point of view. Thus, the maximum value of 7 is copied back to
the root of the tree. The play leading to the board with the score of 7 is
the best choice, provided that the opponent is a good player. This play
forces the opponent into a choice between positions with scores of 8, 7,
and 10. The best that the opponent can do is accept the position with a
score of 7.
To make effective use of the available compute time, Samuel’s pro-
gram adjusts the depth and width of the lookahead search according to
a set of rules (i.e. it uses heuristic search). When a position is unstable, for
example just before a jump, the program looks further ahead. Bad plays
are not explored in depth. Pruning the search in this way affords more
time for evaluation of likely scenarios. To further accelerate processing,
Samuel’s program stores the minimax scores for commonly occurring
board positions. These scores do not need to be recalculated during
execution, as simple table look-up suffices.
In 1962, Samuel’s Checkers playing program was pitted against
Robert Nealey, a blind Checkers master. The computer’s victory was
widely hailed but Nealey wasn’t even a state champion. It would be
thirty years (1994) before a computer program finally defeated the
Checkers world champion.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The AI Winters 91

Samuel retired from IBM in 1966 to take up a research professorship

at Stanford University. He was still programming at the age of eighty-
five when Parkinson’s disease finally forced him to stop working. Samuel
passed away in 1990.
The rudiments of today’s most advanced board game playing algo-
rithms can be seen in Samuel’s 1950s work. Minimax, reinforcement
learning, and self-play are the basis of almost all modern Checkers,
Chess, and Go playing AIs. Moreover, as we shall see, machine learn-
ing has proven to be incredibly effective in tackling complicated data
analysis problems in a great many applications.

The AI Winters
During the late 1950s and 1960s, expectations for AI were sky high.
Supported by Cold War military funding, AI research groups flour-
ished, most notably at MIT, CMU, Stanford University, and Edinburgh
University. In 1958, Newell and Simon predicted that within just ten
years: 83
[A] digital computer will be the world’s Chess champion, unless the rules
bar it from competition.
Four years later, Claude Shannon, founder of information theory, pro-
nounced deadpan to a television camera: 84
I confidently expect that within a matter of ten or fifteen years something
will emerge from the laboratory which is not too far from the robot of
science fiction fame.
In 1968, Marvin Minsky, Head of AI Research at MIT, predicted that: 85
[I]n thirty years we should have machines whose intelligence is compara-
ble to man’s.
Of course, none of these predictions came true.
Why were so many eminent thinkers so spectacularly wrong? The
simplest answer is hubris. These were mathematicians. To them, math-
ematics was the pinnacle of intelligence. If computers could perform
arithmetic, algebra, and logic, then surely more mundane forms of
intelligence must soon yield. What they failed to appreciate was the
variability of the real-world and the complexity of the human brain.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

92 Artificial Intelligence Emerges

Processing images, sounds, and language turned out to be much more

complicated than dealing with equations.
The endemic failure of AI projects caused the funding agencies and
politicians to question the value of this type of research. In the UK,
the British Science Research Council asked Sir James Lighthill, Lucasian
Professor of Mathematics at Cambridge University, to lead a review of
AI research. Published in 1973, his report was damning: 86
In no part of the field have the discoveries made so far produced the major
impact that was then [(around 1960)] promised.
Funding for AI research in the UK was savagely cut. Meanwhile, in the
shadow of the Vietnam War, the Mansfield Amendments of 1969 and
1973 curtailed US governmental spending on research. Only projects
with direct military applications were to be given money.
The first of a succession of ‘AI Winters’ set in. Starved of resources, AI
research groups shrank and shrivelled.
With AI in the doldrums, computer science pivoted to more practical
applications. Given the limitations of computer performance, some
scientists sought to develop fast algorithms to solve important, but
computationally complex, problems. This quest for speed was to lead
to one of the greatest unsolved enigmas in mathematics.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

6
Needles in Haystacks

Der Handlungsreisende – wie er sein soll und was er zu thun hat,

um Aufträge zu erhalten und eines glücklichen Erfolgs in seinen
Geschäten gewiβ zu sein.
Ein alter Commis-Voyageur
Mit einem Titelkupfer, 1832 87

In the 1970s, one of the greatest mysteries in mathematics was un-

covered by an assemblage of researchers investigating the properties of
algorithms. Despite a $1 million prize, that mystery remains unsolved.
At the very heart of the matter is a seemingly innocuous problem.

The Travelling Salesman Problem

The Travelling Salesman Problem asks that the shortest tour of a set of
cities is determined. The cities’ names and the distances between them
are provided. All the cities must be visited once, and only once. The cities
can be visited in any order, as long at the trip starts from, and ends at,
the salesman’s home city. The challenge is to find the tour that offers
the shortest total distance travelled.
The Travelling Salesman Problem was first documented in the 1800s.
At the time, it was a practical concern for commercial travellers jour-
neying between the cities of continental Europe. Later, the Problem was
reformulated as a mathematical plaything by William Hamilton and
Thomas Kirkman.
Let’s say that Berlin is the titular salesman’s home and he must visit
Hamburg, Frankfurt, and Munich (Figure 6.1). The simplest way to find
the shortest tour is by means of an exhaustive search. Exhaustive, or brute-
force, search calculates the length of every possible tour and picks the
shortest. An exhaustive search algorithm proceeds as follows:
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

94 Needles in Haystacks

Figure 6.1 The Travelling Salesman Problem: Find the shortest tour that visits
every city once and returns home.

Take a set of city names as input.

If there is only one city in the set,
then output a tour containing that city alone,
else:
Create an empty list.
Repeat the following for every city in the set:
Create a copy of the set, omitting the selected city.
Apply this algorithm to the reduced set.
Insert the selected city at the start of all the tours
returned.
Append these tours to the list.
Output all tours found.

To start with, the set of all the cities, excluding the home city, is
input to the algorithm. The home city is the known start and end point
of every tour, so it does not need to be included in the search. The
algorithm creates a tree of city visits from the input set (Figure 6.2). The
algorithm relies on two mechanisms. First, it uses repetition—the algo-
rithm selects every city in the input set, one after another, as the next
to be visited. Second, it uses recursion (see Chapter 1). For each city, the
algorithm calls a copy, or instance, of itself. An instance of an algorithm is
another, separate enactment of the algorithm that operates on its own
data. In this case, every instance creates a new sub-tree in the diagram.
After each city is visited, the set of cities input to the next instance
of the algorithm is reduced. Thus, the instances deal with fewer and
fewer cities until there is just one left in the set. When this happens,
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Travelling Salesman Problem 95

Figure 6.2 Tree showing all possible tours. All tours end in Berlin (not shown).

the tree’s leaf instance terminates, returning a tour containing just one
city. The previous instances of the algorithm take this output and add
the selected cities in reverse order. In this way, the algorithm unwinds,
creating tours as it moves up the tree. Once all of the tours have been
traced back to the root of the tree, the original instance of the algorithm
terminates, and the completed list of tours is output.
As the list of tours is being generated, the length of the tours is
calculated by totalling the city-to-city distances.
The operation of the algorithm can be visualized as an animation.
The algorithm constructs the tree from the root. From there, it grows
the topmost path, one city after another, until the top leaf is reached.
It then backtracks one layer and grows the second leaf. Next, it goes back
two layers, before adding the third and fourth leaves. The algorithm
continues sweeping to and fro until the entire tree has been created. In
the end, the algorithm returns to root and terminates.
In the example, Berlin is selected as the home city and so is excluded
from the input set of {Hamburg, Frankfurt, Munich}. The first instance
of the algorithm selects Hamburg, Frankfurt, and Munich in turn as
the first city. For each of these selections, the algorithm spawns a new
instance to explore a sub-tree. After selecting Hamburg as the first city,
the second instance of the algorithm chooses Frankfurt from the set
{Frankfurt, Munich}. It then creates an third instance to deal with the
remaining city: {Munich}. Since there is only one city:
Munich
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

96 Needles in Haystacks

is returned as the only possible tour. The calling instance then prepends
Frankfurt, producing the tour:
Frankfurt, Munich.
The same instance then explores the alternative branch, giving the tour:
Munich, Frankfurt.
These partial tours are returned to the calling instance which adds its
selection, giving the tours:
Hamburg, Frankfurt, Munich;
Hamburg, Munich, Frankfurt.
The sub-trees starting with Frankfurt and Munich are explored in a
similar way. Finally, the complete list of tours is output and the original
algorithm instance terminated.
An exhaustive search such as this is guaranteed to find the shortest
tour. Unfortunately, brute-force is slow. To figure out just how slow
it is, we need to think about the size of the tree. In the example, the
roadmap contains four cities. The cities are fully connected in that every
city is directly connected to all others. Thus, on leaving Berlin, there are
three possible stops. For each of these stops, there are two possible next
cities since the salesman can’t go back home yet or return to the start.
After the first and second stops, there is only one possible third city.
Expanding this out, gives the number of possible tours as 3 × 2 × 1 = 6,
that is, 3 factorial (3!).
Computing the lengths of six tours is a manageable manual com-
putation. What happens if there are 100 cities? One hundred fully
connected cities would give ninety-nine factorial tours, approximately
9 × 10155 (a 9 with 155 zeroes after it). A modern desktop computer
couldn’t possibly cope with that! For the Travelling Salesman Problem,
exhaustive search is surprisingly slow even for roadmaps of seemingly
moderate size.
While more efficient algorithms have been found, the quickest isn’t
much faster than exhaustive search. The only way to significantly speed
up the search is to accept a compromise. You have to accept that the al-
gorithm might not find the shortest possible tour. To date, the best fast
approximation algorithm is only guaranteed to find a path within forty
per cent of the minimum. Of course, compromises and approximations
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Measuring Complexity 97

are not always acceptable. Sometimes, the shortest possible path must
be found.
Over the years, researchers have experimented with programs that
hunt for the shortest tours of real roadmaps. At the beginning of the
computer age (1954) the largest Travelling Salesman Problem with a
known solution contained just forty-nine cities. Fifty years later, the
largest solved tour contained 24,978 Swedish cities. The current state-
of-the-art challenge is a world map of 1,904,711 cities. The shortest tour
found on that map traverses 7,515,772,212 km. Identified in 2013 by Keld
Helsgaun, no one knows if it is the shortest possible tour or not.

Measuring Complexity
The trouble with the Travelling Salesman Problem is the computational
complexity of the algorithm needed to solve it. Computational com-
plexity is the number of basic operations—memory accesses, additions,
or multiplications—required to perform an algorithm. The more oper-
ations an algorithm requires, the longer it takes to compute. The most
telling aspect is how the number of operations grows as the number of
elements in the input increases (Figure 6.3).

Figure 6.3 Graph showing relationships between computational complexity

and the number of inputs to an algorithm.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

98 Needles in Haystacks

A straightforward algorithm has constant complexity. For example,

adding a book to the top of an unsorted pile of novels has a computa-
tional complexity of just one operation. The complexity of adding the
book to the pile is fixed, regardless of the number of books already in
the stack.
Finding a particular title in a bookcase full of books takes more
operations. If the books are unsorted, the librarian might have to check
every single title before the sought-after tome is found. Putting an
additional book in the bookcase increases the worst-case computational
complexity of the search by one operation. In other words, the com-
plexity of title search grows in proportion to the number of books on
the shelves. For this problem, computational complexity is linear with
the number of books.
Algorithms for sorting books have still greater complexity. Insertion
Sort (see Introduction) organizes books one at a time. Putting a book
on the shelf requires that all books already in place are either scanned
past or shifted over. As a consequence, Insertion Sort has computational
complexity proportional to the number of books squared. This gives a
quadratic relationship between the number of books and the number of
operations.
As might be expected, Quicksort (see Introduction) has lower com-
plexity. Quicksort repeatedly splits books into piles based on selected
pivot letters. When the piles contain five or fewer books, Insertion Sort
is applied to every pile and the piles are transferred in order to the
shelf. On average, the complexity of Quicksort is equal to the number
of books multiplied by the logarithm of the number of books. Since
the logarithm of a quantity grows more slowly than the quantity itself,
Quicksort’s complexity is lower than that of Insertion Sort. Quicksort’s
average computational complexity is quasilinear.
Algorithms for adding, searching for, and sorting books have what
is called polynomial computational complexity. A polynomial time al-
gorithm has computational complexity proportional to the number
of inputs to some constant power. In the constant complexity case,
the power is zero. For linear, it is one and for quadratic, the power is
two. Polynomial time algorithms can be slow for very large numbers of
inputs, but in the main, they are tractable on modern computers.
The more challenging algorithms are those with superpolynomial
computational complexity. These methods have complexity in excess
of polynomial time. The number of operations needed to perform
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Complexity Classes 99

superpolynomial time algorithms explodes as the number of inputs

increases. The exhaustive search algorithm for solving the Travelling
Salesman Problem has superpolynomial time complexity. As we have
seen, the number of operations needed is equal to the number of
cities factorial. Adding a single city to a tour multiplies the number
of operations required by the number of cities already on the map. This
multiplicative effect causes an extremely rapid expansion in complexity
as the number of cities grows.
Much work has been conducted on reducing the computational
complexity of such algorithms. The tricks that can be played depend
on the specifics of the problem. Sometimes the structure of the input
data can be exploited to quickly identify obvious solutions or partial so-
lutions. In other cases, using additional data storage enables reductions
in the number of operations. In much the same way, the index of a book
increases the number of pages but greatly reduces the time needed to
find a given keyword.
It is clear that for every problem, there must be a fastest algorithm.
The trick is to find it. In the 1960s and 1970s, a handful of theoreticians
began to investigate the limits of algorithmic complexity. Much of what
we now know about the topic stems from that formative work.

Complexity Classes
Computational problems are graded according to the complexity of
the fastest known algorithms that solve them (Figure 6.4; Table 6.1).
Problems that can be solved with polynomial time algorithms are
referred to as P problems (polynomial time). P problems are considered
quick to solve. For example, sorting is a P problem.

Figure 6.4 Common problem complexity classes.

 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

100 Needles in Haystacks

Table 6.1 Table listing the common complexity classes.

Class Solution Time Verification Time

P polynomial polynomial
NP not specified polynomial
NP\P > polynomial polynomial
>NP > polynomial > polynomial
NP-Complete most complex P polynomial
NP-Hard P convert to NP-Complete not specified

Problems whose solutions can be verified using a polynomial time

algorithm are referred to as NP problems (non-deterministic polynomial
time). How long NP problems take to solve is not defined. Some can be
solved in P time; others cannot. Since solving a problem is a way to verify
a given solution, P problems are by definition also NP problems. In other
words, the set of all P problems is a subset of the set of NP problems.
Problems that are in NP but not in P are called NP\P problems (NP
minus P). These problems are slow to solve but the answers, when they
are known, are quick to check. Sudoku is an NP\P problem.
Sudoku is a Japanese number puzzle played on a 9x9 grid. At the start
of the puzzle, some squares are blank, and some contain digits. The goal
is to fill every blank with a digit in the range 1 to 9. The difficulty is that
a digit may only appear once per row and once per column. A brute-
force approach tries all possible digit placements. As a consequence,
brute-force search is slow. Its complexity is super-polynomial with grid
size. In contrast, verifying a completed grid is quick. The checker need
only scan the rows and columns for blank squares and duplicated digits.
Thus, checking can be completed in polynomial time. Like other NP\P
problems, Sudoku is slow to solve but quick to check.
Beyond this, are greater than NP problems (>NP). These are problems
that cannot be solved or verified in polynomial time. Their solvers
and checkers require superpolynomial time. The Travelling Salesman
Problem is one such problem: solving it requires factorial time. The
only way to check that an answer is the minimum tour is to run the
solver again. Thus, verifying the solution also takes factorial time. Slow
to solve and slow to verify problems are >NP.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Complexity Classes 101

In 1971, Stephen Cook of the University of Toronto, Canada pub-

lished a paper that was to have major ramifications for complexity.
Cook had just joined the University, having been passed over for tenure
at the University of California, Berkeley. 91 Cook’s paper uncovered
deep relationships between certain problem types. The paper led to one
of the greatest mysteries in mathematics, the so-called ‘P versus NP
Problem’, which asks, ‘Can a polynomial time algorithm be found that
will solve all NP problems?’
The importance of the question is clear. A polynomial time algorithm
for solving all NP problems would revolutionize a host of applications.
Previously intractable problems could be quickly solved. For example,
more efficient schedules could be derived in industries ranging from
transportation to manufacturing. Molecular interactions could be pre-
dicted, accelerating drug design and solar panel development.
The most complex NP problems are called the NP-Complete problems.
Cook showed that a polynomial time algorithm for solving an
NP-Complete problem could be used to solve all NP problems in
polynomial time. In other words, a fast algorithm for solving one
NP-Complete problem would, by extension, turn all NP\P problems
into mere P problems. As a result, the set NP would suddenly equal
the set P.
Furthermore, Cook’s work proved that some >NP problems can
be transformed, in polynomial time, into NP-Complete problems. This
transformation is achieved by processing the >NP problem’s inputs in
such a way that an NP-Complete algorithm can finish the computation.
Therefore, finding a fast algorithm for solving an NP-Complete prob-
lem would provide faster solvers for these >NP problems. Collectively,
this group of NP-Complete and >NP problems that are amenable to
P-time transformation to NP-Complete problems are called the NP-
Hard problems. A polynomial time algorithm for solving an NP-Complete
problem would lead to fast solvers for all NP-Hard problems, as well.
The NP-Complete problems are revered. A polynomial time algo-
rithm for solving any one of them would likely win a Fields Medal, the
mathematical equivalent of the Nobel Prize.
The Travelling Salesman Problem is NP-Hard. A simplified version
of the Travelling Salesman Problem, the Travelling Salesman Deci-
sion Problem, is known to be NP-Complete. The problem asks, ‘For a
given roadmap, can a tour be found which is less than a specified target
distance?’ Whereas, verifying a solution to the Travelling Salesman
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

102 Needles in Haystacks

Problem requires solving the problem again (>NP), verifying a given

solution to the decision problem is fast—simply measure the length of
the given tour and compare the result to the specified distance. Thus,
the Decision Problem is slow to solve but fast to verify (NP\P). The set
of NP-Complete problems includes the Knapsack Packing Problem, the
game of Battleships, and the Graph Colouring Problem.
In the year 2000, the Clay Mathematics Institute of Cambridge, Mas-
sachusetts announced $1 million prizes for solutions to seven Mil-
lennium Problems. These Problems were selected as being the most
important in all of mathematics. The ‘P versus NP Problem’ was one
of the seven. The Institute offers the prize to anyone who can provide a
polynomial time algorithm for solving an NP-Complete problem, or a
definitive proof that no such algorithm exists.
Most researchers now think that P is not equal to NP, and never will
be. This conclusion stems from nigh on forty years of failed attempts
to find fast algorithms to solve NP-Complete problems. On the other
hand, a proof that such an algorithm cannot exist remains elusive. So
far, the $1 million Clay Mathematics Institute prize remains unclaimed.
Cook received the ACM Turing Award in 1982. In the citation, he was
lauded for transforming ‘our understanding of the complexity of com-
putation’. Afterwards, Richard Kamp, Professor Emeritus of Electronic
Engineering and Computer Science at the University of California,
Berkeley, ruefully wrote: 91
It is to our everlasting shame that we were unable to persuade the Math
department to give him tenure.

Short Cuts
The Travelling Salesman Problem is one of many combinatorial optimization
problems, which require that a number of fixed elements be combined
in the best way possible. In the case of the Travelling Salesman Problem,
the fixed elements are the city-to-city distances and ‘the best way possi-
ble’ is the shortest tour. The fixed elements can be arranged in a myriad
of ways. The goal is to find the single, best arrangement.
Practical combinatorial optimization problems abound. How best to
allocate staff to tasks in a large factory? What flight schedule maximizes
revenue for an airline? Which taxi should pick up the next customer
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Short Cuts 103

Figure 6.5 Edsger Dijkstra, inventor of the route-finding algorithm, 2002.

(© 2002 Hamilton Richards.)

so as to maximize profits? All of these questions require a fixed set of

resources to be assigned in the best way possible.
As we have seen, brute-force search algorithms try out all possible
combinations and select the best. In 1952, Edsger Dijkstra (Figure 6.5)
came up with a fast algorithm that solves the world’s most common
combinatorial optimization problem. His algorithm is now embedded
in billions of electronic devices.
Dijkstra was Holland’s first professional programmer. Having com-
pleted a programming course in England in 1951, he was offered a part-
time job as a computer programmer in the Mathematisch Centrum
(Mathematical Centre) in Amsterdam. Inconveniently, the Centrum
didn’t have any computers. Short on funding, and these being the early
days of computing, the Centrum was in the process of building one.
Dijkstra worked part-time at the Centrum while studying Mathematics
and Physics at the University of Leiden. After three years, he felt he
couldn’t keep both programming and physics going. He had to choose
one or the other. He loved programming, but was it a respectable
profession for a serious young scientist? Dijkstra went to see Adri-
aan van Wijngaarden, the Director of the Computation Department.
Van Wijngaarden agreed that programming was not a discipline in its
own right, just yet. However, van Wijngaarden confidently predicted
that computers were here to stay and that this was merely the begin-
ning. Could Dijkstra not be one of those who turned programming into
a respectable discipline? By the time he left van Wijngaarden’s office an
hour later, Dijkstra’s path in life was set. He completed his studies in
physics as quickly as he could.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

104 Needles in Haystacks

One year on and Dijkstra was in another predicament. The Centrum

was due to host some important guests. The highlight of their visit was
to be the Centrum’s now operational computer. Dijkstra was asked to
give a demonstration of the machine’s capabilities. Since his guests knew
little about computers, Dijkstra decided that his demo should focus on
a practical application. He hit upon the idea of writing a program to
determine the shortest driving route between two Dutch cities. While
he was satisfied with the concept, a difficulty remained. There was no
fast algorithm for finding the shortest route between two cities.
While out shopping one morning, he and his fiancée stopped at a cafe.
Whilst sipping coffee on the terrace, Dijkstra invented an efficient route-
finding algorithm in about twenty minutes. It would be three years
before he got around to publishing it. It just didn’t seem particularly
important.
Dijkstra’s algorithm is akin to playing a board game. It finds the
shortest route by moving a token between cities on a roadmap. As
the token moves across the map, the cities are annotated with the
route taken and the accumulated distance from the starting point.
When the token leaves a city, the city’s name is crossed off a list of cities
to visit so that the token cannot return.
To begin, the token is placed on the city of origin. The city’s name
and zero distance are recorded next to it. Every directly connected city
is then considered. The distance travelled from the starting point to
these cities is computed. This is achieved by adding the number beside
the token to the length of the link from the token to the directly
connected city. If the city is already annotated with a distance less
than this computed value, the existing annotation is left unchanged.
If the newly calculated value is less than the recorded distance, the
annotation is replaced. The new distance is written down along with
the route taken. The route taken is the list of cities beside the token,
followed by the name of the new city. When these steps have been
carried out for all directly connected cities, the token is forwarded
to the city that has the least annotated distance and that has not yet
been visited. This process—check and update the directly connected
cities, shift the token—is repeated until the token arrives at the desired
destination.
Imagine Dijkstra’s algorithm operating on a Dutch roadmap
(Figure 6.6). Let’s say that the point of departure is Amsterdam and the
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Short Cuts 105

Figure 6.6 The Routing Finding problem requires finding the shortest route
between two cities. This roadmap shows the distances between the principal
Dutch cities in kilometres.

destination is Eindhoven. To start, the token is placed on Amsterdam.

The capital is annotated with:
Amsterdam 0.
The Hague and Utrecht are directly connected to Amsterdam and so
are annotated with:
Amsterdam-The Hague 60;
Amsterdam-Utrecht 50.
The total distance to Utrecht is least, so the token is moved there. The
cities directly connected to Utrecht are then considered. Thus, Tilburg
and Rotterdam are annotated with:
Amsterdam-Utrecht-Tilburg 50+80=130,
Amsterdam-Utrecht-Rotterdam 50+60=110.
The Hague and Amsterdam are not updated since the distances to them
from Amsterdam via Utrecht (90 and 100, respectively) are greater than
the paths already recorded (60 and 0).
The token is transferred to The Hague, the city with the shortest
cumulative distance (60) which has not yet been visited. Going from
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

106 Needles in Haystacks

Amsterdam to Rotterdam via The Hague is a shorter journey than the

previous trip through Utrecht. Hence, Rotterdam is updated with:
Amsterdam-The Hague-Rotterdam 60+30=90.
The token is now moved to Rotterdam. Directly connected Eindhoven
is then updated with:
Amsterdam-The Hague-Rotterdam-Eindhoven 90+110=200.
The path to Tilburg is longer (170 via Rotterdam) than the annotation
already in place (130 via Utrecht), so Tilburg is not updated.
The as-yet unvisited city with the least accumulated distance is
Tilburg, so the token is shifted there. The total distance to Eindhoven
via Tilburg is 165, which is less than the current annotation (200 via
Rotterdam). As a result, Eindhoven’s note is replaced with:
Amsterdam-Utrecht-Tilburg-Eindhown 165.
The token is transferred to Eindhoven and the algorithm is complete.
The shortest path is Amsterdam-Utrecht-Tilburg-Eindhoven with a
total distance of 165 km.
Substituting time for distance allows Dijkstra’s algorithm to find the
fastest route instead of the shortest route.
Dijkstra algorithm is important for two reasons. First, it is guaranteed
to find the shortest route. Second, the algorithm is fast. It prunes the
search on the fly, avoiding bad solutions and concentrating its efforts on
good ones. Third, routing problems are ubiquitous. Every person and
vehicle in the world must navigate.
Dijkstra’s algorithm soon became popular in computing circles. In
1968, it was enhanced by three researchers working at the Stanford
Research Institute (SRI) in California. The researchers were part of
the team that built Shakey the Robot. Shakey was the first general-
purpose mobile robot with reasoning capability. Clunky by today’s stan-
dards, Shakey was essentially a large box containing a small computer
mounted on powered wheels. Its pointy metal ‘head’ supported a large
video camera and an ultrasonic range finder. Preposterous as it now
seems, Life magazine went so far as to call Shakey the ‘first electronic
person’.
Since Shakey was mobile, it needed to be able to navigate. While
incorporating this functionality, the development team spotted an inef-
ficiency in Dijkstra’s algorithm. The algorithm occasionally wastes time
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Stable Marriages 107

moving the token to cities that lead away from the final destination.
These cities look promising as they have short links to the city marked
with the token. However, they lead in the wrong direction and are
ultimately eliminated. To remedy this flaw, Peter Hart, Nils Nilsson,
and Bertram Raphael proposed the A* (a-star) algorithm. A* utilizes a
modified distance metric. In Dijkstra’s original algorithm, the metric
is the distance travelled. In A*, the metric is the distance travelled so
far plus the straight-line distance from the current city to the final
destination. Whereas Dijkstra’s algorithm only considers the path so far,
A* estimates the length of the complete route, from start to finish. As
a result, A* is less inclined to visit cities that take the token away from
the destination.
Nowadays, variants of A* are used in every navigation app on the
planet, from sat navs to smartphones. To allow greater accuracy, cities
have since been replaced by road intersections, but the principles remain
the same. As we will see, derivatives of Dijkstra’s algorithm are now used
to route data over the Internet.
In fulfilment of van Wijngaarden’s prophecy, Dijkstra went on to
make a series of significant contributions to the precarious emerging
discipline that became computer science. Most notably, he originated
algorithms for distributed computing, whereby multiple computers coop-
erate to solve highly computationally complex problems. In recogni-
tion of his work, Dijkstra was the 1972 recipient of the ACM Turing
Award.

Stable Marriages
Roadmaps are not the only source of combinatorial optimization prob-
lems. Matching problems seek to pair items in the best way possible.
A classic matching problem is pairing hopeful college applicants with
available places. The challenge is to assign high school graduates to
courses in a way that is both fair and satisfies as many students and
colleges as possible. As with other combinatorial optimization prob-
lems, matching becomes difficult as the number of items increases. Fast
algorithms are mandatory even for moderate numbers of students.
The seminal paper on matching was published by David Gale and
Lloyd Shapley in 1962. The pair struck up a friendship at Princeton
University, New Jersey, where both studied for PhDs in Mathematics.
After Princeton, Gale joined Brown University in Rhode Island,
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

108 Needles in Haystacks

and Shapley moved to the RAND Corporation. Gale and Shapley’s

paper tackled the age-old problem of matching dating singles for
marriage.
The Stable Marriage Problem seeks to pair men and woman for
marriage. At the outset, female participants rank all males according
to their own preferences (1st, 2nd, 3rd, and so on), and, vice versa, the
men rank the women. The objective is to pair the participants in such
a way that the set of marriages is stable. The set is considered stable if
there exists no male and female who prefer each other to their spouses.
Such a circumstance could lead to divorce.
Gale and Shapley’s paper proposed a remarkably simple algorithm
for solving the Stable Marriage Problem. So simple, in fact, that the
authors had trouble getting it published at all. In the paper, Gale and
Shapley used the Stable Marriage Problem as a proxy for a range of
real-world two-way matching problems. Two-way refers to the fact that
both parties state preferences, not just one party. The Gale–Shapley
algorithm quickly became the de facto two-way matching method. The
algorithm is still employed in a wide variety of applications, including
matching critically ill patients with organ donors.
The Gale–Shapley algorithm matches men and women over suc-
cessive rounds. In each round, all of the unattached men make one
marriage proposal. (For simplicity, the paper assumes that the partic-
ipants are heterosexual and that the marriage proposals are made by
the men. In fact, the proposals could equally be made by the women—
it doesn’t matter to the algorithm.) Every unattached man proposes
to the woman who has his highest preference and has not previously
rejected him. If the woman receiving the proposal is not engaged,
she automatically accepts the proposal. If she is already engaged, she
considers her preferences for her new suitor and her existing fiancée. If
she prefers her new suitor, she spurns her fiancée and is betrothed to
her new paramour. If she prefers her existing fiancée, she turns down
the advances of the would-be interloper. Once her decision is made,
the algorithm continues to the next unmarried man. When all of the
unmarried men have made proposals, the algorithm moves on to the
next round. At this stage, a rebuffed male fiancée is free to make a fresh
proposal of marriage to someone else. The algorithm ends when all
participants are engaged for marriage.
Consider six friends—three men and three woman—living across
the hall from each another in two New York apartments close to
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Stable Marriages 109

Table 6.2 Table showing marriage preferences.

Preference Alex Ben Carlos Diana Emma Fiona

1st Diana Diana Diana Ben Ben Carlos
2nd Emma Fiona Fiona Alex Alex Alex
3rd Fiona Emma Emma Carlos Carlos Ben

Central Park. For the sake of discretion, let’s call them Alex, Ben, Carlos,
Diana, Emma, and Fiona. All know one another other. All are single. All
have matrimony on their minds. When asked, they state the marriage
preferences listed in Table 6.2.
In the first round, Alex proposes to Diana. Diana accepts because
she is currently available. Next, Ben proposes to the popular Diana.
Since Diana prefers Ben to Alex, she ditches the latter and accepts Ben’s
proposal. Carlos also proposes to Diana and gets turned down flat. In
the second round, Alex and Carlos are unattached. Alex proposes to
Emma, number two on his list. Emma is, as yet, without a partner,
so she acquiesces to Alex’s advances. Carlos requests Fiona’s hand in
marriage and she agrees since she is unattached. That’s it. Everyone
is now engaged—Alex & Emma, Ben & Diana, Carlos & Fiona. All of
the marriages are stable. Emma would prefer Ben but she can’t have
him since he would rather couple with Diana, his soon-to-be wife.
Similarly, Alex and Carlos have unrequited feelings for Diana but she
is set to marry the man of her dreams, Ben. Fiona is happily betrothed
to Carlos—her number-one pick—even though he, also, has a crush
on Diana.
As a dating scheme, the Gale–Shapley algorithm is brutal—all those
rejections! Nevertheless, one wonders if humans seeking partners in-
tuitively follow a procedure akin to the Gale–Shapley algorithm. In
real life, explicit proposals and firm replies become surreptitious smiles,
longing gazes, inquiries via friends, and polite refusals. While there are
similarities, there are differences. In truth, preferences evolve over time.
The emotional cost of separation means that individuals are reluctant
to make drastic changes. Despite these contrasts, some online dating
agencies now use the Gale–Shapley algorithm to match clients.
One of the biggest matching exercises in the world is carried out
every year by the National Residency Matching Program (NRMP). The
program pairs medical school graduates with internship opportunities
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

110 Needles in Haystacks

in hospitals across the US. Currently, the program matches 42,000

graduate applicants with 30,000 hospital vacancies per annum. When
it was established in 1952, the NRMP adopted a matching algorithm
from a previous clearing house. The Boston Pool algorithm was used
by the NRMP for forty years. During the 1970s, it came to light that the
Boston Pool algorithm was, in fact, the same as the independently de-
veloped Gale–Shapley method. Embarrassingly, the unknown Boston
Pool team had beaten the eminent mathematician–economist duo to
the punch by more than a decade. Of course, their academic paper
included a formal proof, whereas the Boston Pool algorithm was ad hoc.
In the 1990s, the noted economist and mathematician Alvin Roth was
engaged to revamp the NRMP matching algorithm. Moving with the
times, Roth’s new method allows medical couples to seek co-location.
It also seeks to prevent rogue applicants gaming the system to their
advantage. Unlike the Boston Pool, Roth’s technique relies on one-way
matching, whereby only the applicants’ preferences are considered.
Shapley and Roth were jointly awarded the Nobel Prize in Economics
in 2012 for their work on game theory. This is the branch of mathematics
concerned with competition and cooperation between intelligent deci-
sion makers. Shapley, the elder of the two, is widely seen as the grand
theoretician who laid the groundwork for Roth’s practical studies on
how markets operate. One of the highlights of their Nobel Prize citation
was the Gale–Shapley algorithm. Gale passed away in 2008 and so was
ineligible for the Prize. Shapley died in 2016, at the grand old age of 92.
Roth continues to work at Stanford and Harvard Universities.

Artificial Evolution
In the 1960s, John Holland (Figure 6.7) took a radical approach to
solving combinatorial optimization problems. Uniquely, his algorithm
was four billion years old!
Holland was born in Fort Wayne, Indiana in 1929. Like Dijkstra, Hol-
land caught the computer programming bug while studying physics.
At MIT, he wrote a program for the Whirlwind computer. Funded by
the US Navy and Air Force, Whirlwind was the first real-time computer
to incorporate an on-screen display. The machine was designed to
process radar data and provide early warning of incoming aircraft and
missiles. After a brief sojourn programming at IBM, Holland moved to
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Artificial Evolution 111

the University of Michigan to pursue a Master’s degree followed by a

PhD in Communication Sciences. It would be several years before the
term ‘computer science’ came into vogue. Holland’s PhD supervisor
was Arthur Burks, the man who ran the ENIAC demo for the press in
1946.
While browsing in the university library, Holland came upon an old
book by Ronald Fisher called The Genetical Theory of Natural Selection (1930).
The book employed mathematics to investigate natural evolution. Hol-
land later recollected: 103
That was the first time I realized that you could do significant mathemat-
ics on evolution. The idea appealed to me tremendously.
Inspired, Holland determined that he would replicate evolution in
a computer. He pursued this singular idea throughout the 1960s and
1970s. Convinced he was on to something significant, Holland authored
a book detailing his findings in 1975. Sales were disappointing. The
research community wasn’t much interested.
Almost twenty years later, Holland penned an article on the topic
of genetic algorithms for the popular science magazine Scientific American.
A second edition of his book went to press that same year. Finally,
genetic algorithms broke through to the mainstream of computing
research. Holland’s long-neglected text has now been cited (formally
referenced) in more than 60,000 books and scientific papers—a box
office hit by academic standards.

Figure 6.7 Designer of the first genetic algorithms, John Holland. (© Santa Fe
Institute.)
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

112 Needles in Haystacks

Natural evolution adapts a species to its environment by means of

three mechanisms—selection, inheritance, and mutation. The process acts on
a species living in the wild. While the individuals in a species share
many characteristics, there is variation from one animal to the next.
Some traits are beneficial for survival, while others are detrimental.
Selection refers to the fact that individuals with beneficial characteristics
are more likely to survive to adulthood and reproduce. Inheritance
is the tendency of children to display similar physical traits as their
parents. Thus, children of survivors tend to have the same beneficial
characteristics. Mutation is the manifestation of random alterations
in the genetic material passed on to a child. Depending on what part of
the chromosome is affected, a mutation might have no effect, a limited
impact, or cause an extreme change. Over many generations, selection
and inheritance mean that a population will tend to have a greater
proportion of individuals with traits that are beneficial to survival and
reproduction. Mutation is the joker in the pack. Mostly, it has no
impact on the population. Sometimes, it plants the seed for a radical
beneficial change.
The classic example of natural evolution at work is the pepper moth.
The name comes from the insect’s mottled wings which look like
black pepper sprinkled on white paper. The insect’s appearance is a
form of camouflage, which makes it difficult for predatory birds to
spot the creatures against the bark of the local trees. In eighteenth-
century England, peppered moths were predominantly pale in colour.
Strangely, by the end of the nineteenth century, almost all peppered
moths in the major cities were dark. Over the course of a century, the
English urban pepper moth population had changed colour.
It became apparent that the change was precipitated by the Industrial
Revolution. Rapid exploitation of fossil fuels had led to the construc-
tion of large numbers of factories belching out great plumes of smoke
and soot. Tree bark, walls, and lampposts gradually turned black. Pale
moths became more vulnerable to predators and were killed in greater
numbers. The dark moths thrived, passing their natural camouflage
on to their children. Over time, the balance in numbers within the
population shifted in favour of the dark moths.
Holland believed that artificial evolution could be used to solve com-
binatorial optimization problems. His idea was that possible solutions
could be thought of as individuals within a population. Drawing on
genetics, he suggested that every solution be represented by a sequence
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Artificial Evolution 113

of letters. For example, to solve the Travelling Salesman Problem, a

tour might be represented using the first letter of the cities: BFHM.
He considered this sequence analogous to a living organism’s chromo-
some (DNA).
Holland’s genetic algorithms act on this pool of artificial chromosomes.
Selection is performed by evaluating every artificial chromosome and
discarding the worst performing ones. Inheritance is mimicked by in-
termingling letter sequences to create the next generation of chromo-
somes. Mutation is replicated by randomly replacing letters in a handful
of chromosomes. These three processes are repeated to produce gener-
ations of chromosomes until, finally, an acceptable solution is identified
among the population.
Over many generations, the three mechanisms act in concert to
increase the proportion of good solutions in the population. Selection
guides the search towards better solutions. Inheritance mixes promising
answers in unexpected ways to produce new candidate solutions. Muta-
tion raises the diversity in the population, opening up new possibilities.
Holland’s algorithm can be summarized as follows:

Generate a population of chromosomes at random.

Repeat the following steps:
Evaluate every chromosome’s performance.
Discard the worst performing chromosomes.
Pair the surviving chromosomes at random.
Mate every pair to produce two children chromosomes.
Add the children to the population.
Randomly alter a small number of chromosomes.
Stop repeating after a fixed number of generations.
Output the best performing chromosome.

In one of his books, the renowned biologist Richard Dawkins re-

ported using a genetic algorithm to reveal a secret message. Dawkins
used chromosomes directly as guesses for the secret message. He per-
formed selection by comparing each chromosome to the secret mes-
sage. Dawkins’ fitness function is unclear but most likely his program
calculated a score for each chromosome. Perhaps the correct letter in
the correct place was given a score of +2 points, the correct letter in
the wrong place was worth +1, and an incorrect letter was worth 0.
The chromosomes with the lowest scores were discarded. The highest
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

114 Needles in Haystacks

scoring chromosomes were ‘mated’ by means of crossover. Crossover

generates chromosomes for offspring by exchanging fragments of the
parents’ chromosomes. A random position is selected within the chro-
mosome. The first child’s chromosome is a copy of the father’s chromo-
some up to that point and a copy of the mother’s thereafter. The second
child’s chromosome is a copy of the mother’s up to the same point
and a duplicate of the father’s beyond it. For example, if the parents’
chromosomes are:
ABCDEF, LMNOPQ,
and the crossover point is the third letter, the children’s letter sequences
are:
ABCOPQ, LMNDEF.
Dawkins ran his genetic algorithm in a computer. Initially, the chro-
mosomes were random letter sequences. After ten generations, he
reported that the highest scoring chromosome in the population was:
MDLDMNLS ITPSWHRZREZ MECS P.
After twenty, it was:
MELDINLS IT ISWPRKE Z WECSEL.
After thirty:
METHINGS IT ISWLIKE B WECSEL.
After forty-one generations, the algorithm found the secret message:
METHINKS IT IS LIKE A WEASAL.
Unusually, Dawkins’ example evaluates the chromosome directly,
comparing the letters to the secret message. To solve real-world prob-
lems, the chromosome normally controls the construction of a solu-
tion, which is then evaluated. For example, Holland used chromosomes
to control the behaviour of computer programs simulating commodity
markets. Even though they were not initially programmed to do so,
his agents evolved so as to contrive speculative bubbles and financial
crashes.
NASA employed genetic algorithms to design a radio antenna for
one of its space missions. This time, the chromosomes controlled the
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Artificial Evolution 115

shape of the antenna. The fitness of the chromosomes was evaluated by

calculating the sensitivity of the antenna to incoming radio signals. An
evolved X-band antenna flew on the NASA Space Technology 5 mission
in 2006.
In 1967, Holland was appointed Professor of Computer Science and
Engineering at the University of Michigan. As well as inventing genetic
algorithms, he made major contributions to complexity and chaos
theory. Very unusually, he also became a Professor of Psychology.
Holland passed away in 2015 at the age of 86. David Krakauer, the
President of the Santa Fe Institute, said of him: 112
John is rather unique in that he took ideas from evolutionary biology
in order to transform search and optimization in computer science,
and then he took what he discovered in computer science and allowed
us to rethink evolutionary dynamics. This kind of rigorous translation
between two communities of thought is a characteristic of very deep
minds.
While genetic algorithms remain popular, they are often not the
most efficient way to solve a given optimization problem. They work
best when there is little understanding of how good solutions are put
together. Driven by random permutations, genetic algorithms blindly
search the design space. Since they are easy to program, researchers
often just let the computer do the work with a genetic algorithm,
rather than spend precious time inventing a new fast search algorithm.
While Holland was working on his genetic algorithms in Michigan,
the seeds of a revolution in computing were being sown by an obscure
agency within the US Department of Defense. Spurred by the Cold War
and guided by two visionaries, America began to interlink its computer
networks. The endeavour was to have far reaching implications. Along
the way, it sparked a new alchemy—the fusion of algorithms with
electronics.
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

7
The Internet

It seems reasonable to envision […] a ‘thinking centre’ that will

incorporate the functions of present-day libraries.
The picture readily enlarges itself into a network of such centres,
connected to one another by wide-band communication lines
and to individual users by leased-wire services. In such a system,
the speed of the computers would be balanced, and the cost of
the gigantic memories and the sophisticated programs would be
divided by the number of users.
JCR Licklider
Man-Computer Symbiosis, 1960 113

On 4 October 1957, the Soviet Union launched the world’s first artificial
satellite into Earth orbit. Sputnik 1 was a radio transmitter wrapped
up in a metal sphere just 60 cm in diameter. Four radio antennas were
fixed to the globe’s midriff. For twenty-one days, Sputnik broadcast a
distinctive and persistent ‘beep-beep-beep-beep’ signal. The signal was
picked up by radio receivers worldwide as the satellite passed overhead.
Sputnik was a sensation. Suddenly, space was the new frontier and the
Soviet Union had stolen a march on the West.
US President Dwight D. Eisenhower resolved that the US should
never again take second place in the race for technological supremacy.
To this end, Eisenhower established two new governmental agencies. He
charged the National Aeronautics and Space Administration (NASA)
with the exploration and peaceful exploitation of space. Its malevo-
lent sibling—the Advanced Research Projects Agency (ARPA)—was
instructed to fund the development of breakthrough military tech-
nologies. ARPA was to be the intermediary between the military and
research-performing organizations.
The Cold War beckoned.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

118 The Internet

ARPANET
Nineteen sixty-two saw the founding of ARPA’s Information Processing
Techniques Office (IPTO). The IPTO’s remit was to fund research and de-
velopment in information technology (i.e. in computers and software).
The Office’s inaugural director was JCR (Joseph Carl Robnett) Licklider
(Figure 7.1).
Licklider (born 1915) hailed from St. Louis. Universally liked, every-
one called him ‘Lick’ for short. Lick never lost his Missouri accent. At
college, he studied an unusual cocktail of physics, mathematics, and
psychology. He graduated with a PhD in Psychoacoustics, the science
of the human perception of sound, from the University of Rochester.
He then worked at Harvard University before joining MIT as an As-
sociate Professor. It was at MIT that Licklider first became interested
in computers. It transpired that he had a talent, even a genius, for
technical problem solving. He carried his new-found passion into his
role at ARPA.
Licklider wrote a series of far-sighted papers proposing new computer
technologies. Man-Computer Symbiosis (1960) suggested that computers
should work more interactively with users, responding to their requests
in real-time rather than by batch print-out. He proposed the creation
of an Intergalactic Computer Network (1963) enabling integrated operation of
multiple computers over great distances. In Libraries of the Future (1965),
he argued that paper books should be replaced by electronic devices
that receive, display, and process information. He jointly published a
paper in 1968 that envisaged using networked computers as person-to-
person communication devices. In a ten-year burst of creativity, Lick-
lider predicted personal computers, the Internet, eBooks, and email. His
imagination far outpaced reality. His writings set out grand visions for
others to chase.
The first step was Project MAC (Multiple Access Computing) at MIT.
Up to that point computers had a single user. Project MAC constructed
a system whereby a single mainframe computer could be shared by
up to thirty users working simultaneously. Each user had their own
dedicated terminal consisting of a keyboard and screen. The computer
switched its attention between the users, giving each the illusion that
they had a single, but less powerful, machine at their disposal.
Two years after Licklider’s departure from ARPA, Bob Taylor was
appointed Director of the IPTO (1965). From Dallas, Taylor (born 1932)
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

ARPANET 119

Figure 7.1 JCR Licklider, computer network visionary. (Courtesy MIT Museum.)

had previously worked for NASA. Like Licklider, Taylor’s background

was in psychology and mathematics. Unusually, for a senior execu-
tive in a research funding body, Taylor didn’t hold a doctoral degree.
Whereas Licklider was an ideas guy, Taylor had an almost uncanny
knack of delivering breakthrough technologies. Taylor was to spend the
next thirty years turning Licklider’s visions into reality.
Taylor’s enthusiasm for computer networking was borne out of
simple frustration. Three computer terminals sat in his office at the Pen-
tagon. Each terminal was connected to a different remote computer—
one at MIT, another in Santa Monica, and the third at Berkeley. The
computers were not interconnected in any way. To pass a message from
one machine to another, Taylor had to type it in again on the other
terminal. He remembers: 119
I said, oh man, it’s obvious what to do. If you have these three terminals,
there ought to one terminal that goes anywhere you want to go.
At Taylor’s behest, IPTO programme manager Larry Roberts compiled
a Request For Quotation for construction of a computer network. The
new network was to be called the ARPANET. Initially, the ARPANET
would link four sites, with a possible expansion to thirty-five. The
winning bid was submitted by Bolt, Beranek, and Newman Technolo-
gies (BBN) of Cambridge, Massachusetts.
Three months after the Apollo 11 moon landing, on 29 October 1969,
Charley Kline sent the first message on the ARPANET. Kline was a
student programmer in Leonard Kleinrock’s group at the University
of California, Los Angeles (UCLA). His intent was to send a ‘LOGIN’
command to a computer at SRI, 400 miles away. However, the system
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

120 The Internet

crashed after the second character was received. As a result, the first
message sent on the ARPANET was the inauspicious fragment ‘LO’.
About an hour later, after a system re-start, Kline tried again. This time,
the login worked.
ARPANET was one of the first networks to employ packet-switching.
The technique was invented independently by Paul Baran and Donald
Davies. Baran, a Polish–American electrical engineer, published the idea
in 1964 while working for the RAND Corporation. Davies, a veteran of
Turing’s ACE project, developed similar ideas while working at the NPL
in London. It was Davies that coined the terms ‘packet’ and ‘packet-
switching’ to describe his algorithm. Davies was later part of the team
that built the world’s first packet-switched network—the small-scale
Mark I NPL Network in 1966.
Packet-switching (Figure 7.2) solves the problem of efficient transport
of messages across a network of computers. Imagine a network of
nine computers that are physically interconnected by means of cables
carrying electronic signals. To reduce infrastructure costs, each com-
puter is connected to a small number of others. Directly connected
computers are called neighbours, regardless of how far apart they really
are. Sending a message to a neighbour is straightforward. The message
is encoded as a series of electronic pulses transmitted via the cable to
the receiver. In contrast, sending a message to a computer on the other
side of the network is complicated. The message must be relayed by
the computers in between. Thus, the computers on the network must
cooperate to provide a network-wide communication service.
Before packet-switching, communication networks relied on dedi-
cated end-to-end connections. This approach was common in wired
telephony networks. Let’s say that computer 1 wishes to communicate
with computer 9 (Figure 7.2). In the conventional circuit-switching scheme,
the network sets up a dedicated electrical connection from computer
1 to computer 3, from 3 to 7, and from 7 to 9. For the duration of
the message exchange, all other computers are blocked from sending
messages on these links. Since computers typically send short sporadic
messages, establishing dedicated end-to-end connections in this way
makes poor use of network resources. In contrast, packet-switching
provides efficient use of network links by obviating the need for end-
to-end path reservation.
In packet-switching, a single message is broken up into segments.
Each segment is placed in a packet. The packets are independently
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

ARPANET 121

Figure 7.2 Packet-switching on a small network.

routed across the network. When all of the packets have been received
at the destination, a copy of the original message is assembled. The
network transfers packets from source to destination in a series of hops.
At every hop, the packet is transmitted over a single link between two
computers. This means that a packet only blocks one link at a time.
Thus, packets from different messages can be interleaved—one after
another—on a single link. There is no need to reserve an entire end-
to-end connection.
The downside is that links can become congested. This happens when
incoming packets destined for a busy outgoing link must be delayed and
queued.
A packet-switched data network is similar to a road network. The
packets comprising a single message are akin to a group of cars making
their way to a destination. The cars don’t need the entire path be
reserved in advance. They simply slot in when the road is free. They may
even take different paths to the destination. Each car wends its own way
through the network as best it can.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

122 The Internet

A packet contains a header, a portion of the message (the payload), and

a trailer (Figure 7.2). The header consists of a start of packet marker, the
unique ID of the destination, a message ID, and a sequence number. The
sequence number is the packet count within the message (1, 2, 3, …).
The sequence numbers are used by the destination computer to assem-
ble the payloads in the correct order. The trailer contains error checking
information and an end-of-packet marker.
Computers on the network continuously monitor their connections
for incoming packets. When a packet is received, it is processed by the
computer. First of all, the ID of the packet’s destination is read from
the header. If the receiving computer is not the final destination, the
device re-transmits the packet on the link offering the fastest route
to that destination. The link to be used is determined by means of a
routing table. The routing table lists the IDs of all computers, or groups
of computers, on the network. For each destination, it records the
outgoing link that provides the fastest path to that machine. The
receiving computer forwards the packet to this link. If the receiving
computer is the intended destination, the message ID and sequence
number are inspected. When all of the packets in a single message have
arrived, the computer concatenates their payloads by sequence number
so as to recover the original message (see the Appendix for the full
algorithm).
Packet-switching is an example of a distributed algorithm. Distributed
algorithms run on multiple, independent, but co-operating, comput-
ers. Every computer performs its own individual task, which con-
tributes in some way to the grand scheme.
An essential component of a packet-switching system is the algo-
rithm used to populate the routing table. The routing algorithm deter-
mines how best to populate the routing table. It is the routing algorithm
that decides which packets go where.
The original ARPANET routing algorithm based its decisions on
exchange of path delay information. In addition to a routing table, every
computer maintains a delay table (Figure 7.3). The delay table lists every
computer ID on the network and, for each machine, the estimated
time taken to deliver a packet to it from that point. Periodically, all
computers send their delay table to all of their neighbours. On receipt of
a neighbour’s delay table, a computer adds the time it would take to get
a packet to the neighbour. The updated table thus contains a list of every
computer ID and the time it would take to get a packet there via the
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

ARPANET 123

Figure 7.3 The routing table for computer 3 listing the packet destination, the
delay via computers 1, 2, 4, 7, and 8, and the ID for the neighbour giving the
shortest route from computer 3 to the destination. Delay is measured in hops.

neighbour. The computer compares these times with those already in

the routing table. The current entry is updated if the new path is faster.
In summary:

For every neighbouring computer:

Get its delay table.
For every destination in the table:
Add the delay from this computer to the neighbour.
If the new delay is shorter than the delay in the routing
table, then:
Store the link ID and the new delay in the routing
table.
The delay of a path can be measured as the number of hops to
the destination. Alternatively, the number of packets queued along
the path can be counted. This latter option provides an estimate of the
congestion on the path and typically yields better results.
One of the benefits of packet-switched networks is robustness to
failure. If a link fails, or becomes overly congested, computers can
detect the resulting increase in packet queue length and adjust their
routing tables so as to avoid the bottleneck. The disadvantage of packet-
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

124 The Internet

switching is that packet delivery times are unpredictable. Packet-

switching is a ‘best effort’ service.
After two years at ARPA, the networking visionary, Licklider joined
IBM. Later he returned to MIT. Following a second term as IPTO
Director, Lick lider again resumed his teaching and research career at
MIT. He finally retired in 1985.
Bob Taylor went on to found Xerox’s Palo Alto Research Center’s
(PARC’s) Computer Science Laboratory in 1970. During his tenure,
PARC built a succession of remarkable real-world prototypes. The
Center developed a low-cost, packet-switched networking technol-
ogy for interconnecting computers within a single building. That
technology—Ethernet—is now the most common wired computer
networking standard in the world. The Lab also invented the point-and-
click graphical user interface, which was to become ubiquitous in Apple’s
Mac computers and in the Microsoft Windows operating system. In
one of the greatest corporate miscalculations of the twentieth century,
Xerox failed to capitalize on PARC’s breakthroughs. Taylor left PARC
in 1983, later establishing a research centre for Digital Equipment
Corporation. He retired in 1996 and passed away in 2017.
The original four-node packet-switched ARPANET grew slowly but
steadily with, on average, one node (computer site) being added per
month. The ARPANET was not the world’s first, nor only, computer
network. Nevertheless, it was to be the progenitor of the largest com-
puter network on the planet: the Internet.

Internetworking
The first public demonstration of the ARPANET took place at the
International Computer Communication Conference (ICCC) in Wash-
ington in October 1972. Organized by Bob (Robert) Kahn of BBN,
the demo connected forty plus computers. Hundreds of previously
sceptical industry insiders saw the demo and were impressed. Packet-
switching suddenly seemed like a good idea after all.
Kahn was from New York (born 1938). He graduated with MA and
PhD degrees from Princeton before joining BBN. Shortly after the ICCC
demo, he moved to the IPTO to oversee further development of the
network. Kahn reckoned that the next big challenge was not adding
more computers to the ARPANET, but connecting the ARPANET to
other networks.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Internetworking 125

The ARPANET was all a bit samey—the nodes and links used similar
fixed line technologies. Kahn imagined a hyperconnected world in
which messages would travel transparently between devices on all kinds
of networks—wired, radio, satellite, international, mobile, fixed, fast,
slow, simple, and complex. This was a glorious concept, to be sure.
The question was how to make it work? Kahn came up with a concept
he called open-architecture networking. The approach seemed promising but
the devil was in the details. In 1973, Kahn visited the Stanford lab of
ARPANET researcher Vint Cerf and announced: 126
I have a problem.
Vint (Vinton) Cerf was born in New Haven, Connecticut in 1943. He
graduated Stanford University with a BS degree in Mathematics before
joining IBM. A few years later, he chose to enrol in graduate school at
UCLA. It was at UCLA that Cerf began to work on the ARPANET. Here
too he met Kahn for the first time. Shortly after Kahn’s ICCC demo,
Cerf returned to Stanford as a professor.
Working together, Kahn and Cerf (Figure 7.4) published an outline
solution to the inter-networking problem in 1974. In it, they proposed
an overarching protocol that would be common to all networked com-
puters. The protocol defined a set of messages and associated behaviours
that would allow computers to communicate across technologically

Figure 7.4 Inventors of TCP/IP: Vint Cerf (left, 2008) and Robert Kahn (right,
2013). (Left: Courtesy Vint Cerf. Right: By Вени Марковски | Veni Markovski -
Own work / CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=
26207416.)
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

126 The Internet

disparate networks. A protocol is an agreed sequence of messages that

allows computers to interact. For example, in the human world, we
use ‘Hello’ to initiate a conversation and ‘Goodbye’ to finish it. Ev-
eryone knows what the standard messages mean and what should
happen next.
As they incorporated feedback from other networking aficionados,
Cerf and Kahn’s outline snowballed into a detailed technical specifi-
cation. That specification was to become the first description of the
Transmission Control Protocol / Internet Protocol (TCP/IP). Cerf and
Kahn believed that TCP/IP would solve the problem of interconnecting
distinct computer networks.
The first trial of TCP/IP came on 27 August 1976. A computer in-
stalled in an SRI delivery truck, affectionately nicknamed ‘the bread
van’, parked south of San Francisco communicated with a fixed line
ARPANET node in the University of Southern California. The demon-
stration proved that TCP/IP allowed message exchange between radio
and wired networks.
The following year, communication across three networks was
tested. This time, nodes in Norway and the United Kingdom were
internetworked with sites in the US via a transatlantic satellite link.
To Cerf, this three-network test was the real thing—proper inter-
networking using TCP/IP. The experiment on 22 November 1977
heralded the dawn of the Internet.
Today, TCP/IP is the protocol by which computers transmit and
receive messages on the Internet. To the user, the most visible element
of TCP/IP is the global naming convention for computer nodes. A com-
puter’s Internet Protocol address uniquely identifies it on the network.
An IP address consists of four numbers separated by dots. For example,
the IP address of Google’s search page is 172.217.11.174. Later, textual
names for nodes were added for convenience. The domain name system
converts these textual names to numeric IP addresses. Thus, the domain
name google.com translates to the IP address 172.217.11.174.
In 1983, the ARPANET’s original communications protocol—NCP—
was replaced by TCP/IP. Around the same time, the ARPANET’s routing
algorithms were upgraded. Dijkstra’s route-finding algorithm began to
be used for packet-routing (see Chapter 6). Ironically, the algorithm for
finding the shortest path between cities is now far more frequently used
to determine the fastest route for data packets traversing the Internet.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Fixing Errors 127

ARPANET was formally decommissioned in 1990. By then, it was

just another network within a global collection of interconnected net-
works. The backbone of the Internet in the US was now NSFNET.
ARPANET’s demise went largely unnoticed. Licklider, the original com-
puter networking visionary, passed away that same year.
Ex-BBN-employee Bob Kahn went on to found the Corporation for
National Research Initiatives (CNRI) in 1986. CNRI conducts, funds, and
supports research on forward-looking information technologies. He
and Vint Cerf founded the Internet Society in 1992 to promote TCP/IP.
Cerf continued his work on the Internet, accumulating a lengthy list
of corporate and not-for-profit appointments. Perhaps the most eye-
catching is his Chief Internet Evangelist role at Google. NASA also
engaged Cerf as a consultant on its plans for the Interplanetary Internet.
Cerf and Kahn were the recipients of the ACM Turing Award for 2004.
TCP/IP enabled the interworking of a global patchwork of distinct
computer networks. There is no central controller on the Internet.
So long as computers adhere to the TCP/IP protocols and naming
conventions, they can be added to the network. In the period 2005 to
2018, the Internet had 3.9 billion users—more than half of the world’s
population. This figure looks set to increase further.
Thanks to TCP/IP, humanity has never been so interconnected.

Fixing Errors
Communication systems such as the Internet are designed to transfer
exact copies of messages from the transmitter to the receiver. To achieve
this, the data in the packet is converted to an electronic signal that
is sent from the transmitter to the receiver. The destination device
converts the received signal back into data. Often, the received signal
is contaminated by electronic noise. Noise is the general term for any
unwanted signal that corrupts the intended signal. Noise can arise from
natural sources or interference from nearby electronic equipment. If
the noise is strong enough relative to the signal, it can lead to errors
in conversion of the signal back into data. Obviously, errors are not
desirable. An error in a sentence might be tolerated. However, an error
in your bank balance would not be, unless it happened to be in your
favour! For this reason, communications systems incorporate error
detection and correction algorithms.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

128 The Internet

The simplest way to check for, and correct, an error is repetition.

To ensure correct delivery, an important piece of data might be sent
three times in succession. The receiver compares all three copies. If
they match, then it can be assumed that no error occurred. If just two
copies match, then the odd-one-out can be assumed to be in-error. The
two matching copies are taken to be correct. If no copies match, then
the true message is unknown and retransmission must be requested.
For example, if the received error protected message is
HELNO, HFLLO, HELLP
then it is probable that the original was HELLO since the three Hs match,
two Es out-vote one F, the next three Ls match, two Ls beat one N, and
two Os out-vote one P.
Repetition works well. However, it is very inefficient. If every packet
is sent three times, then the number of packets that can be transmitted
per second is a third of what it was before error protection was added.
Checksums are much more efficient. The idea is that all characters
(letters, digits, and punctuation) are converted to numbers. These num-
bers are added together and the total—the checksum—is transmitted
together with the message. On receiving the packet, the computer
recalculates the checksum and compares the result of this computation
with the checksum included in the packet. If the calculated and received
checksums match, then, most likely, the received packet is error-free.
Of course, it is possible that two equal, and opposite, errors occurred
in the data or that the checksum and data suffer exactly the same
errors. However, these eventualities are extremely unlikely. Most of
the time, a checksum mismatch indicates that a transmission error
has occurred.
For example, converting HELLO to integers gives:
8 5 12 12 15.
Adding these values produces a checksum of 52. The packet sent is then:
8 5 12 12 15 - 52.
If transmission is error-free, the receiver obtains an exact copy of the
message. It recalculates the checksum as 52. This matches the checksum
in the packet, and all is well.
What if the first character is mistakenly received as an F?
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Fixing Errors 129

6 5 12 12 15 - 52.
This time, the calculated checksum (50) does not match the checksum
at the end of the packet (52). The receiver knows that an error has
occurred.
Checksums are common. For example, the International Standard
Book Number (ISBN) at the front of this book contains a checksum. All
books printed after 1970 have an ISBN that uniquely identities the title.
Current ISBNs are thirteen digits long and the last digit is a check digit.
The check digit allows computers to verify that the ISBN has been typed
in or scanned correctly.
A basic checksum merely detects errors. It is impossible to work out
which number is incorrect. The error could even be in the checksum it-
self. Ironically, in this case, the message itself is correct. Basic checksums
require retransmission of the message to fix an error.
Richard Hamming (Figure 7.5) wondered if checksums could be
modified to provide error correction, as well as detection.
Born in Chicago in 1915, Hamming studied Mathematics, ultimately
attaining the PhD degree from the University of Illinois. As the Sec-
ond World War ended, Hamming joined the Manhattan Project in Los
Alamos. He worked, as he put it, as a ‘computer janitor’, running
calculations on IBM programmable calculators for the nuclear physi-
cists.130 Disillusioned, he moved to Bell Telephone Laboratories in New
Jersey. Bell Labs was the research wing of the burgeoning Bell Telephone
Company founded by Alexander Graham Bell, the inventor of the
telephone. In the late 1940s and 1950s, Bell Labs employed a stellar cast
of communications researchers. Hamming was in his element: 130

Figure 7.5 Richard Hamming, inventor of error-correcting codes, 1938.

(Courtesy of the University of Illinois Archives, 0008175.tif.)
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

130 The Internet

We were first-class troublemakers. We did unconventional things in un-

conventional ways and still got valuable results. Thus, management had
to tolerate us and let us alone a lot of the time.
Renowned for his bad jokes, not everyone appreciated Hamming’s way
of doing things: 130
He is very hard to work with because he does a lot of broadcasting and
not a lot of listening.
Hamming’s interest in error correction stemmed from his own frus-
trations in dealing with an unreliable relay-based computer. All too
often, he would leave a program running over the weekend only to
discover on the Monday morning that the job had failed due to a fault
in the computer. Exasperated, Hamming wondered why the machine
couldn’t be programmed to find and fix its own errors.
The simplest form of checksum is the parity bit. Modern electronic
computers process information as binary numbers. Unlike decimal
numbers, binary numbers only use two digits: zero and one. Whereas
in decimal every column is worth ten times the one to its right, in
binary every column is worth double. Thus, in binary—moving right-
to-left from the fractional point—we have the units, twos, fours, eights,
sixteens, and so on. For example:
1011 binary = (1 × 8) + (0 × 4) + (1 × 2) + (1 × 1) = 11 decimal.
Similarly, counting from zero to fifteen in binary gives the sequence:
0, 1, 10, 11, 100, 101, 110, 111,
1000, 1001, 1010, 1011, 1100, 1101, 1110, 1111.
Thus, four binary digits, or bits, can represent the decimal integers from
zero to fifteen.
A parity bit can be appended to a binary number for the purposes of
error detection. The value of the parity bit (0 or 1) is selected such that
the total number of 1 bits, including the parity bits, is even. For example,
the data word:
010001
is protected by appending a parity bit with the value zero:
0 1 0 0 0 1 - 0.
This keeps the number of 1 bits even (i.e. two 1-bits).
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Fixing Errors 131

As with checksums, the parity bit is sent along with the data word, that
is, the sequence of data bits. To check for errors, the receiver simply
counts the number of 1s. If the final count is even, then it can be as-
sumed that no error occurred. If the count is odd, then, in all likelihood,
one of the bits suffered an error. A 0 has been mistakenly flipped to a 1,
or vice versa. For example, an error in bit two gives the word:
0 0 0 0 0 1 - 0.
This time, there is an odd number of 1 bits, indicating that an error has
occurred.
In this way, a single parity bit allows detection of a single error. If two
bits are in error, then the packet appears valid, but is not. For example:
1 0 0001-0
seems to be correct since there is an even number of 1s. As a conse-
quence, additional parity bits are needed when there is a high error rate.
Hamming devised a clever way to use multiple parity bits to detect
and correct single bit errors. In Hamming’s scheme, every parity bit
protects half of the bits in the word. The trick is that no two parity bits
protect the same data bits. In this way, every data bit is protected by a
unique combination of parity bits. Hence, if an error occurs, its location
can be determined by looking at which parity bits are affected. There can
only be one data bit protected by all of the parity bits showing errors.
Let’s say that a data word containing eleven data bits is to be
transmitted:
1 0 1 0 1 0 1 0 1 0 1.
In Hamming’s scheme, eleven data bits require four parity bits. The
parity bits, whose value is to be determined, are inserted into the data
word at positions which are powers of two (1, 2, 4, and 8). Thus, the
protected word becomes:
? ? 1 ? 0 1 0 ? 1 0 1 0 1 0 1,
where the question marks indicate the future positions of the parity bits.
The first parity bit is calculated over the bits at odd numbered positions
(numbers 1, 3, 5, etc.). As before, the value of the parity bit is selected to
ensure that there is an even number of 1s within the group. Thus, the
first parity bit is set to 1:
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

132 The Internet

1 ? 1 ? 0 1 0 ? 1 0 1 0 1 0 1 .
The circles indicate the bits in the parity group. The second parity bit is
calculated over the bits whose positions, when written in binary, have a
1 in the twos column (2, 3, 6, 7, etc.):
1 0 1 ?0 1 0 ?1 0 1 01 0 1 .
The third parity bit is calculated over bits whose positions, in binary,
have a 1 in the fours column (4, 5, 6, 7, 12, etc.):
101 1 0 1 0 ?101 0 1 0 1 .
The fourth parity bit is calculated over bits whose positions, in binary,
have a 1 in the eights column (8, 9, 10, 11, etc.):
1011010 0 1 0 1 0 1 0 1 .
This then is the final protected data word, ready for transmission.
Now, imagine that the protected data word suffers an error at bit
position three:
1 0 0 1 0 1 0 0 1 0 1 0 1 0 1.
The receiver checks the word by counting the number of ones in the
four parity groups:
1 0 0 1 0 1 0 0 1 0 1 0 1 0 1 = 5 ones;
1 0 0 10 1 0 01 0 1 01 0 1 = 3 ones;
100 1 0 1 0 0101 0 1 0 1 = 4 ones;
1001010 0 1 0 1 0 1 0 1 = 4 ones.
The first and second parity groups both show errors (i.e. they have an
odd numbers of ones). In contrast, the third and fourth groups do not
indicate errors (i.e. they have an even numbers of ones). The only data
bit that is in the first and second groups and is not in the third and fourth
groups is bit three. Therefore, the error must be in bit number three.
The error is easily corrected by flipping its value from 0 to 1.
Hamming’s ingenious algorithm allows detection and correction of
single errors at the cost of a small increase in the total number of bits
sent. In the example, four parity bits protect eleven data bits—just a
thirty-six per cent increase in the number of bits. Hamming codes are
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Secret Messages 133

remarkably simple to generate and check. This makes them ideal for
high-speed processing, as required in computer networks, memory, and
storage systems. Modern communication networks employ a mixture
of Hamming codes, basic checksums, and newer, more complex, error-
correcting codes to ensure the integrity of data transfer. A mistake in
your bank balance is extremely unlikely.
After fifteen years in Bell Labs, Hamming returned to teaching,
taking up a position at the Naval Postgraduate School in Monterey,
California. Hamming received the Turing Award in 1968 for his codes
and other work on numerical analysis. He died in 1998 in Monterey, just
one month after finally retiring.
One of the great flaws of the Internet is that it was not designed
with security in mind. Security has had to be grafted on afterwards,
with mixed results. One of the difficulties is that packets can be easily
read en route by eavesdroppers using electronic devices. Encryption cir-
cumvents eavesdropping by altering a message in such a way that only
the intended recipient can recover the original text. An eavesdropper
might still intercept the altered text, but the scrambled message will be
meaningless.
Until the end of the twentieth century, encryption algorithms were
intended for scenarios in which the encryption method could be agreed
in absolute secrecy before communication took place. One imagines a
queen furtively passing a top-secret codebook to a spy at a clandestine
rendezvous. However, this approach doesn’t translate well to computer
networks. How can two computers secretly exchange a codebook when
all data must be sent over a vulnerable public network? At first, it
seemed that encryption wasn’t at all practical in the brave new world
of the computer network.

Secret Messages
Encryption was used in ancient Mesopotamia, Egypt, Greece, and India.
In most instances, the motivation was secure transmission of military
or political secrets. Julius Caesar employed encryption for important
personal letters. The Caesar Cipher replaces every letter in the original
text with a substitute letter. The substitute letter is a fixed number
of places away from the original in the alphabet. To make patterns
more difficult to spot, Caesar’s Cipher strips away spaces and changes all
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

134 The Internet

letters to capitals. For example, a right shift by one place in the alphabet
leads to the following encryption:
Hail Caesar
IBJMDBFTBR.
The As become Bs, the Es change to Fs, and so on. Any Zs would be
replaced by As since the shift wraps around the end of the alphabet.
The encrypted message—the ciphertext—is sent to the receiver. The
receiver recovers the original message—the plaintext—by shifting every
letter one place left in the alphabet. The Bs become As and so on
returning the original ‘HAILCAESAR’ message. Thanks to the patterns
in natural language, the missing spaces are surprisingly easy to infer.
Traditional encryption methods, such as the Caesar Cipher, rely on
an algorithm and a secret key. The key is a piece of information that is
essential for successful encryption and decryption. In the Caesar Cipher,
the key is the shift. The algorithm and the key must be known to the
sender and the intended recipient. Typically, security is maintained by
keeping the key secret.
No encryption scheme is perfect. Given sufficient time and a clever
attack, most codes can be broken. The Caesar Cipher can be attacked by
means of frequency analysis. An attacker counts the number of times
that each letter of the alphabet occurs in the ciphertext. The most
commonly occurring is probably the substitute for the vowel E, since
E is the most common letter in the English language. Once a single
shift is known, the entire message can be decrypted. Almost all codes
have vulnerabilities. The question is, ‘How long does the attack take to
perform?’ If the attack takes an unacceptably long period of time, then
the cipher is secure from a practical point of view.
In computer networks, key distribution is problematic. The only
convenient way to pass a key is via the network. However, the network is
not secure against eavesdropping. Sending a secret key via the Internet
is equivalent to making it public. How could a sender and a receiver
possibility agree on a secure key if all they can do is send public messages?
The question became known as the Key Distribution Problem. The first
glimmer of a solution came from a group working at Stanford in the
early 1970s.
Martin Hellman was born in New York in 1945. He studied Electrical
Engineering at New York University before moving to California to
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Secret Messages 135

study for his MSc and PhD degrees at Stanford. After stints at IBM and
MIT, Hellman returned to Stanford in 1971 as an Assistant Professor.
Against the advice of his peers, Hellman started to work on the Key
Distribution Problem. Most thought it foolhardy to expect to find
something radically new—something that the well-resourced US Na-
tional Security Agency (NSA) had missed. Hellman was unperturbed.
He wanted to do things differently to everyone else. In 1974, Hellman
was joined in the hunt for a solution by Whitfield Diffie.
From Washington, DC, Diffie (born 1944) held a degree in Mathe-
matics from MIT. After graduation, Diffie worked programming jobs
at MITRE Corporation and his alma mater. However, he was capti-
vated by cryptography. He struck out to conduct his own independent
researches on key distribution. On a visit to IBM’s Thomas J. Watson
Laboratory in upstate New York, he heard about a guy called Hellman
who was working on similar stuff at Stanford. Diffie drove 5,000 miles
across the US to meet the man who shared his passion. A half-hour
afternoon meet-up extended long into the night. A bond was formed.
The duo was joined by PhD student Ralph Merkle. Born in 1952,
Merkle had previously come up with an innovative approach to the
Key Distribution Problem while studying as an undergraduate at the
University of California, Berkeley.
In 1976, Diffie and Hellman published a paper describing one of the
first practical algorithms for public key exchange. The paper was to
revolutionize cryptography. The myth that all keys had to be private
was shattered. A new form of coding was born: public key cryptography.
The Diffie–Hellman–Merkle key exchange scheme showed that two
parties could establish a secret key by means of public messages. There
was a hitch, though. Their method required the exchange and process-
ing of multiple messages. As a result, the algorithm was not ideal for use
on networks. However, their paper did suggest an alternative.
Traditional encryption algorithms use a symmetric key, meaning that
the same key is used for encryption and decryption. The drawback with
symmetric encryption is that the key must be keep secret at all times.
This requirement creates the Key Distribution Problem.
In contrast, public key encryption uses two keys: an encryption key
and a different—asymmetric—decryption key. The pair of keys must meet
two requirements. First, they must work successfully as an encryption–
decryption pair, i.e. encryption with one and decryption with the other
must return a copy of the original message. Second, it must be impossi-
ble to determine the decryption key from the encryption key. Therein
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

136 The Internet

lies the beauty of public key cryptography. If the decryption key cannot
be determined from the encryption key, then the encryption key can be
made public. Only the decryption key needs to be keep secret. Anyone
can use the public encryption key to send a secret message to the private
key holder. Only the recipient who possesses the private decryption key
can decipher and read the message.
Imagine that Alice wants to be able to receive encrypted messages
(Figure 7.6). She creates an asymmetric key pair by means of a key
generation algorithm. She keeps the decryption key to herself. She
publicly releases the encryption key on the Internet. Let’s say that Bob
wants to send a secret message to Alice. He obtains Alice’s encryption
key from her Internet posting. He encrypts the message using Alice’s
encryption key and sends the resulting ciphertext to Alice. On receipt,
Alice decrypts the ciphertext using her private decryption key. In short:

Alice generates the encryption and decryption key pair.

Alice keeps the decryption key to herself.
She makes the encryption key public.
Bob encrypts his message using Alice’s public encryption key.
Bob sends the encrypted message to Alice.
Alice decrypts the encrypted message using her secret
decryption key.

The scheme works beautifully with one proviso. There must be

no way to determine the private decryption key from the public en-
cryption key. Therein lay the difficulty. No one knew how to create
asymmetric keys where the decryption key could not be worked out
from the encryption key. What was needed was a one-way function—a
computation whose input could not be easily inferred from its output.
If such a function could be found, its output could be the basis of the
public key and its input the basis of the private key. There would be no
way to reverse the computation. Attackers would be unable to recover
the decryption key.
Diffie and Hellman’s paper described public key encryption but
did not offer a one-way function. The concept was inspired, but they
couldn’t find a way to make it work.
At MIT’s Laboratory for Computer Science in Boston, Ronald Rivest
read Diffie and Hellman’s paper with mounting excitement. There and
then he undertook to hunt for a suitable one-way function. A function
that would unlock public key encryption. Rivest persuaded two friends
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Secret Messages 137

Figure 7.6 Public key cryptography.

and colleagues, Adi Shamir and Leonard Adleman, to help him in the
search. All three held Bachelor’s degrees in Mathematics and PhDs in
Computer Science. Rivest (born 1947) hailed from New York state, Adi
Shamir (1952) was a native of Tel Aviv, Israel, and Adleman (1945) grew
up in San Francisco. The impromptu team spent a year generating
promising ideas for one-way functions only to discard each and every
one. None were truly one-way. Perhaps there was no such thing as a
one-way function.
The trio spent Passover 1977 as guests at a friend’s house. When Rivest
returned home, he was unable to sleep. Restless, his mind turned to the
one-way encryption problem. After a while, he hit upon a new function
that might just work. He wrote the whole thing down before dawn.
The next day, he asked Adleman to find a flaw in his scheme—just
as Adleman had done for every other suggestion. Curiously, Adleman
couldn’t find a weakness. The method seemed to be robust to attack.
They had found a one-way function. Rivest, Shamir, and Adleman’s
algorithm for key generation was published later that same year. It
quickly became known as the RSA algorithm from its inventors’ initials.
RSA is now the cornerstone of encryption on the Internet.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

138 The Internet

RSA encryption and decryption are reasonably straightforward. The

encryption key consists of two numbers—a modulus and an encryption
exponent. The decryption key also contains two numbers—the same
modulus and a unique decryption exponent.
To begin, the original textual message is converted to a sequence of
numbers. Encryption is applied to groups of numbers as follows:

Calculate the input number to the power of the encryption

exponent.
Calculate the remainder after dividing this number by the
modulus.
Output the remainder.

To decipher:

Calculate the number received to the power of the decryption

exponent.
Calculate the remainder after dividing by the modulus.
Output the remainder.

Let’s say that the encryption key is (33, 7), the message is 4, and the
decryption key is (33, 3). Calculating 4 to the power of 7 (that is, 4
multiplied by itself 7 times) gives 16,384. The remainder after dividing
16,384 by 33 is 16. So, 16 is the ciphertext.
To decipher, calculate 16 to the power of 3 giving 4,096. The remain-
der after dividing 4,096 by 33 is 4. The output, 4, is the original message.
How does the scheme work? The process hinges on clock arithmetic.
You may have come across the number line—an imaginary line with the
integers marked off evenly along it, much like a ruler. Starting at zero,
the number line stretches off into infinity. Now imagine that there are
only 33 numbers on the line (0–32). Roll up this shortened line into a
circle. The circle looks much like the face of an old-fashioned wall clock
marked with the numbers 0 to 32.
Imagine starting at zero and counting around the clock. Eventually,
you come to 32 and after that the count goes back to 0, 1, 2 and so on.
You keep going around and around the clock face.
Clock arithmetic mirrors the effect of the remainder operation. Di-
viding 34 by 33 gives a remainder of 1. This is the same as going around
the clock once and taking one more step.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Secret Messages 139

In the example, encryption moves the clock hand 16,384 steps around
the clock face. In the end, the clock hand is left pointing to 16: the
ciphertext. Decryption starts at 0 and moves the clock hand 4,096 steps
around the clock face. In the end, the clock points to 4: the original
message.
The encryption and decryption keys are complementary. The key
pair is especially selected so that one exponent undoes the effect of the
other. The number of complete rotations around the clock face does
not matter. In the end, all that matters is the digit that the hand is
pointing to.
The key pair is produced by means of the RSA key generation al-
gorithm. This is the heart of the RSA encryption. The first two steps
contain the one-way function:

Pick two large prime numbers with similar values.

Multiply them to get the modulus.
Subtract one from each prime number.
Multiply the results to give the totient.
Choose the encryption exponent as a prime number between 1
and the totient.
Repeat the following steps:
Choose a constant value.
Calculate the constant multiplied by the totient and add
one.
Calculate this number divided by the encryption exponent.
Stop repeating when the result is a whole number.
Let the decryption exponent equal the whole number.
Output the encryption key (modulus and encryption
exponent).
Output the decryption key (modulus and decryption
exponent).

The algorithm is complicated, but let’s try an example. Say we start

with the prime numbers 3 and 11. These are too small to be secure from
attack, but they will do for now. The modulus is then 3 × 11 = 33. The
totient is (3−1)×(11−1) = 20. We select 7 as the encryption exponent
since it is a prime number between 1 and the totient. Taking a constant
value of 1, we have 1 + 1 × 20 = 21. The decryption exponent is then
equal to 21 divided by 7, that is, 3. Thus, the encryption key is (33, 7) and
the decryption key is (33, 3).
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

140 The Internet

Attacking the key pair boils down to finding the two prime numbers
that were multiplied together to give the modulus. Multiplication
disguises the selected primes. For large numbers, there are many pairs
of prime numbers that might have been multiplied together to give the
modulus. An attacker would have to test a huge number of primes to
crack the code. When a large modulus is used, brute-force search for the
original prime numbers is prohibitively time-consuming.
The other steps in the key generation algorithm ensure that the
encryption and decryption are reciprocal. That is, decryption undoes
encryption for all values between zero and the modulus.
A Scientific American article introduced the RSA algorithm to a general
readership in 1977. The item concluded with a $100 challenge to crack
an RSA ciphertext given the encryption key. The modulus was reason-
ably large—129 decimal digits. It took seventeen years to crack the code.
The winning team comprised six hundred volunteers leveraging spare
time on computers all over the world. The plaintext turned out to be
the distinctly uninspiring: 136
The Magic Words are Squeamish Ossifrage
An ossifrage is a bearded vulture. The $100 prize worked out at 16 cents
per person. The point was proven. RSA is military-grade encryption.
Public key encryption is now built into the Secure Socket Layer of
the World Wide Web. When a web site address is preceded by “https:”,
your computer is using SLL and, with it, the RSA algorithm to com-
municate with the remote server. At last count, seventy per cent of
web site traffic used SSL. Over the years, the length of keys has had to
increase to prevent ciphertexts being cracked by the latest supercom-
puters. Today, keys containing 2,048 or more bits (617 decimal digits)
are commonplace.
The academics had proven the naysayers wrong. Governmental elec-
tronic espionage agencies could be beaten at their own game. At least,
that’s how it seemed.
Amidst the hullabaloo over RSA, the Head of the US National Se-
curity Agency stated publicly that the intelligence community had
known about public key encryption all along. This raised eyebrows.
Was his claim factual, bravado, or just plain bluff? Diffie’s curiosity was
piqued. He made discrete enquiries and was directed to look across the
Atlantic towards GCHQ. Government Communication Headquarters
is the UK’s electronic intelligence and security agency. During the
Second World War, it was the body that oversaw the code-breaking work
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Secret Messages 141

at Bletchley Park. Diffie wrangled two names from his contacts: Clifford
Cocks and James Ellis. In 1982, Diffie arranged to meet Ellis in a pub in
Cheltenham. A GCHQ stalwart to the last, the only hint that Ellis gave
was the cryptic: 137
You did a lot more with it than we did.
In 1997, the truth came out. GCHQ published the papers of Cocks,
Ellis, and another actor, Malcolm Williamson. Among the documents
was a history of events at GCHQ written by Ellis. In it, he refers to Diffie–
Hellman–Merkle’s ‘rediscovery’ of public key cryptography.
In the early 1970s, James Ellis hit upon the idea of public key cryp-
tography. His name for the technique was ‘non-secret encryption’. An
engineer by profession, Ellis couldn’t come up with a satisfactory one-
way function. Clifford Cocks happened to hear about Ellis’s discovery
over a cup of tea with Nick Patterson. Cocks, an Oxford and Cambridge
educated mathematician, found himself at a loose end that evening. He
decided to study the one-way encryption problem. Spectacularly, Cocks
solved the problem that evening. Just like the RSA team, he settled on
multiplication of two large prime numbers. This was four years ahead of
Rivest, Shamir, and Adleman. Cocks circulated the idea in the form of
an internal GCHQ paper. Malcolm Williamson picked up on the memo
and added a missing piece on key exchange some months later.
The codebreakers at GCHQ couldn’t find a flaw in Ellis, Cocks, and
Williamson’s unconventional method. Nevertheless, the hierarchy re-
mained unconvinced. Non-secret encryption languished in the office
drawers at GCHQ. Gagged by the Official Secrets Act, the authors
said nothing. They watched on the sidelines as the Stanford and MIT
teams reaped the glory. Ellis never gained the satisfaction of public
recognition. He passed away just a few weeks before the embargo on
his papers was lifted.
Rivest, Shamir, and Adleman won the ACM Turing Award for 2002.
Diffie and Hellman were honoured as the recipients of the 2015 award.
Meanwhile, the Internet grew and grew. By 1985, there were 2,000
hosts (computer sites) on the Internet. Most were owned by academic
institutions. While data transport worked well, the networking pro-
grams of the 1980s were unappealing. Their user interface was text-
heavy, monochrome, and cumbersome to use. To reach a wider audi-
ence, computing needed a makeover.
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

8
Googling the Web

A memex is a device in which an individual stores all books,

records, and communications, and which is mechanised so that
it may be consulted with exceeding speed and flexibility.
[The essential feature of the memex] is a provision whereby any
item may be caused at will to select immediately and automati-
cally another.
Vannevar Bush
The Atlantic, 1945 140

By the 1970s, minicomputers were well established in scientific insti-

tutes, universities, and large companies. Similar in height and girth to
an American-style refrigerator, a minicomputer was far cheaper than
the antiquated mainframe, but was still a hefty purchase. While effective
for large-scale data processing, the machines were not user-friendly.
User terminals consisted of a monochrome monitor and a chunky
keyboard. A fixed grid of green letters and numbers illuminated a black
background. An arcane textual command was required for every action.
In 1976, two California kids—Steve Jobs and Steve Wozniak—
launched the first pre-assembled microcomputer onto an indifferent
market. For the first time, a commercial computer was small enough to
fit on a desktop. Even better, the Apple I was cheap enough that it could
be bought by, and used by, just one person. Apple followed up with an
improved offering the next year. Despite the compact dimensions of
the hardware, sales were muted until the arrival of VisiCalc.
VisiCalc was the world’s first commercial spreadsheet program. Visi-
Calc allowed users to enter text, numbers, and formulae into on-screen
tables. The secret sauce was that VisiCalc would automatically perform
the calculations specified in the formulae when numbers were entered
into the spreadsheet. There was no need to write a program to get a
few calculations done. Suddenly, business users could play with their
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

144 Googling the Web

sales figures without resorting to the IT Department. People bought the

Apple II just to use VisiCalc. The desktop computer market burgeoned.
IBM spotted the disruption late. Digital Equipment Corporation
(DEC) and IBM were the leading suppliers of minicomputers. In a
catch-up play, IBM launched their own Personal Computer in 1981.
Supported by the corporation’s vast sales network, the IBM PC was a
commercial success.
Three years later, Apple beat IBM to the punch once more. Ap-
ple released the first graphical user interface (GUI) for an affordable
computer. Apple’s attractively designed Macintosh computer shipped
with a keyboard, high-resolution screen, and—radically—a mouse.
The mouse allowed users to control the computer by clicking on icons
and menus. Programs now ran side by side in adjustable windows. The
mouse was a hit. Old-fashioned textual commands were for geeks.
Although Apple commercialized the GUI, the technology was in-
vented elsewhere. The mouse came from Douglas Engelbart at SRI. The
GUI was created at Xerox PARC under Bob Taylor’s watch. Steve Jobs
happened to see a demo of PARC’s GUI and had to have one for his own
products.
The Macintosh left Apple’s competitors scrambling. IBM’s software
partner hastily added a GUI into their next operating system, Microsoft
Windows. At the same time, IBM’s PC sales came under pressure from
low-cost copycat, or clone, hardware manufacturers. IBM and Apple
simply couldn’t compete on price. Soon, cheap PC clones running
Windows were on every desktop in every enterprise.
While the GUI made working with a single computer much easier,
accessing data via networks remained a chore. The Internet provided
global connectivity for specialist computer centres. However, network-
ing programs still used textual commands. Even worse, every remote
access system (bulletin boards, library catalogues, remote logins, etc.)
had its own peculiar set of commands. Using the Internet was incredibly
frustrating. Easy-to-use software for sharing data between computers
was sorely needed. Surprisingly, the fix didn’t come from the computer
industry. It came from a European particle physics laboratory.

World Wide Web

Tim Berners-Lee was born in London in 1955. He graduated Oxford
University with a degree in Physics. Following in the footsteps of
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

World Wide Web 145

his parents—who had been programmers on the Ferranti Mark

I—Berners-Lee became a professional software developer. Between
industry jobs, he worked as a contractor at CERN for six months in
1980. Four years later, he returned to work on the design of CERN’s
computer network.
CERN is the European Organisation for Nuclear Research. The body
is based at a sprawling campus in Geneva, Switzerland. In 1984, Berners-
Lee was one of 10,000 staff, students, and visiting researchers working
on a myriad of loosely connected projects. The place was a hodgepodge
of organizational structures, vested interests, cultures, and languages.
Coordination and communication were next to impossible. It seemed
to Berners-Lee that networked computers could assist in the day-to-day
task of information sharing.
Berners-Lee proposed a scheme whereby a user’s desktop computer
could download and view electronic pages stored on remote, or server,
computers. Each page would be held as a data file which could be
transferred over the Internet. The data file would include text and
special formatting tags specifying how the page was to be displayed.
A piece of software called a browser running on the user’s computer
would request the remote page from the server and display it. Each
page would be identified by a unique name consisting of the server’s ID
followed by the file name. Later, a prefix indicating which protocol to
use was added. The complete identifier is now called the page’s uniform
resource locator (URL).
A key feature of Berners-Lee’s proposal was that the pages would
contain hyperlinks. A hyperlink, or link or short, is a piece of text, or an
image, that is tagged with a reference to another web page. When a
user selects the link, their browser automatically displays the referenced
page. Links greatly simplified navigation between pages. Click on a link
and the associated page appears.
Hyperlinks seemed startlingly modern. In fact, the idea wasn’t new
at all. Hyperlinks were first suggested by Vannevar Bush in a visions-
of-the-future article published back in 1945. What was new was that
networked computers and software could turn Bush’s speculation into
reality.
Berners-Lee and Robert Cailliau put together a detailed document
describing the system. Emphasizing the global reach of the Internet
and the hyperlinked connectivity of pages, they named the proposed
system the WorldWideWeb (WWW). Spaces were later added to make
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

146 Googling the Web

the name more readable. The document described two elements:

a formal definition of the file format (i.e. the allowed contents) of
web pages; and a protocol specifying the messages and behaviours that the
browser and server software should use to communicate.
Within a year of receiving approval for the project, Berners-Lee
finished the software for the first WWW browser and server. The world’s
first web site went live on 6 August 1991. The original pages are still
available at:
info.cern.ch/hypertext/WWW/TheProject.html
At Berners-Lee’s insistence, CERN released the WWW specifications
and software for free. Uptake was sluggish. In 1993, a team from the
University of Illinois at Urbana-Champaign led by Marc Andreessen
released a new web browser. Mosaic was compatible with Berners-Lee’s
server software but, crucially, ran on Microsoft Windows. Windows-
based PCs were far more common than the obscure workstations that
Berners-Lee had programmed at CERN. By the end of the year, 500 web
servers were online. Sites on particle physics and computing were joined
by pages devoted to financial news, web comics, movie databases, maps,
webcams, magazines, company brochures, restaurant advertisements,
and pornography.
The following year, as the WWW gained traction, Tim Berners-Lee
left CERN to found and lead the World Wide Web Consortium (W3C).
The W3C operates as a not-for-profit body working in collaboration
with industry partners to develop and promote the WWW. The Con-
sortium is the guardian of the World Wide Web standards to this day.
True to Berners-Lee’s ideals, the WWW standards have remained open,
free, and publicly available. Anyone can build a compatible web browser
or server without having to seek permission or pay a royalty.
The WWW provided the world with a low cost, robust, and easy-to-
use platform for sharing information. It was up to website developers
to figure out what to do with it.

Amazon Recommends
In the year that Berners-Lee left CERN, a Wall Street investment banker
happened upon a startling statistic. Spurred by the popularity of the
Mosaic browser, web usage had grown 2,300 per cent year-on-year.
The number was ludicrous. Double digit growth is hard to come by,
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Amazon Recommends 147

nevermind four digit. The banker found lots of information on the web
but almost nothing for sale. Surely, this was an untapped market. The
question was: ‘What to sell?’
At the time, the Internet was way too slow to stream music or
videos. Product delivery would have to be by the US postal service. An
online store would be like a mail-order business—only better. Cus-
tomers could view an up-to-date product catalogue and place orders
via the web. The banker looked up a list of the top twenty mail-order
businesses. He concluded that book retail would be a perfect fit. The
investment banker had stumbled upon the opportunity of a lifetime.
At thirty years old, Bezos was DE Shaw & Company’s youngest ever
Senior Vice President (VP). Born in Albuquerque, New Mexico, Bezos
grew up in Texas and Florida. He attended Princeton University, grad-
uating in Computer Science and Electrical Engineering. After college,
he worked a series of computer and finance jobs, quickly climbing the
ladder to Senior VP.
Bezos’s brainwave left him with a dilemma. Should he pack in his
six-figure New York banking job to go selling books? Bezos called on
an algorithm that he employed for making life altering decisions like
this: 143
I wanted to project myself forward to age 80 and say, ‘OK, I’m looking
back on my life. I want to minimize the number of regrets I [will] have.’
I knew that when I was 80, I was not going to regret having tried this.
I knew the one thing I might regret is not ever having tried. [That] would
haunt me every day.
Bezos quit his Wall Street job and embarked on a mission to build a
bookstore in a place that didn’t really exist—online.
He needed two things to get his Internet business off the ground: staff
with computer skills and books to sell. Seattle on the northwest coast
of the US had both. The city was home to Microsoft and the country’s
largest book distributors. He and his wife of one year—MacKenzie Bezos
(née Tuttle)—boarded a plane to Texas. On arrival, they borrowed a car
from Bezos’ dad and drove the rest of the way to Seattle. MacKenzie
took the steering wheel while Bezos typed up his business plan on a
laptop computer.
Using his parents’ life savings as seed capital, Bezos set up shop in a
small two-bedroom Seattle house. On 16 July 1995, the amazon.com
website went live.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

148 Googling the Web

Figure 8.1 Greg Linden, designer of the first Amazon recommender system,
2000. (Courtesy Greg Linden.)

Sales were healthy. Amazon re-located to 2nd Avenue. The offices

were so over-crowded that Greg Linden (Figure 8.1), one of the new
hires, was forced to work in the kitchen.
Linden was on a time-out from the University of Washington. He
loved the buzz of the start-up but fully intended to return to college
to do a PhD in Computer Science. In the meantime, Linden reckoned
that he could help Amazon sell more books. He was convinced that a
recommender system would help turn Amazon home page views into book
purchases.
A product recommender system analyses customers’ purchasing
decisions and suggests products that they might like to buy. A reader
that has previously bought a collection of crime novels might well be
tempted into buying a Sherlock Holmes special edition or a Raymond
Chandler novel. Presenting these books as suggestions to the user might
well encourage them to buy one.
In essence, recommendation is advertising. The distinction is that
recommendation is personalized. It is tailored to the individual according
their interests. Linden’s hope was that personalized recommendations
would increase the number of sales per home page view in comparison
to traditional one-size-fits-all advertisements. A handful of experimen-
tal recommender engines were already online. In contrast, Amazon’s
recommender would have to work at scale in a commercial setting.
Linden pitched his idea to management and was given the go-ahead to
build Amazon’s first recommender system.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Amazon Recommends 149

Linden based his algorithm on a simple intuition. If a pair of products

are commonly bought together then it is likely that a customer who
already owns one will buy the other. The items don’t have to have
been originally purchased as a pair. All that matters is that users often
buy both. The reason why doesn’t matter, either. The books could be
by same author. They might be from the same genre. Perhaps they
are both study guides for the same state examinations. In terms of
recommending products, the timing of, and rationale, for the pairing
are irrelevant. All that matters is that customers often buy both items.
Linden’s algorithm records all of the purchases made on the Amazon
website. When a user checks out, the algorithm notes the purchaser’s
unique ID and the names of the books purchased. The algorithm
recovers a list of all previous purchases made by that user. It then pairs
the new items with all of the user’s previous purchases.
Imagine that Mary buys Charlotte’s Web and has previously purchased
The Little Prince and Pinocchio. This creates two new book pairings:
Charlotte’s Web & The Little Prince
Charlotte’s Web & Pinocchio
This information is used to update a product similarity table. The table lists
all of the books on the Amazon web site down the rows and across the
columns. Thus, every pair of books has two entries in the body of the
table. The entries record the number of times that a pair of books (row
and column) have been purchased by a single user. This number is the
similarity score for the pair.
Mary’s purchase of Charlotte’s Web generates two new pairs. As a result,
the entries for Charlotte’s Web and The Little Prince are incremented by one,
as are the entries for Charlotte’s Web and Pinocchio (Table 8.1).
When a user arrives at the Amazon web site, the algorithm uses
the similarity table to generate recommendations. The method begins
by retrieving the user’s purchasing history. For every book that the
user has bought, the algorithm looks up the corresponding row in the
similarity table. The algorithm scans across these rows looking for non-
zero entries. For each, the algorithm looks up the book title at the top
of column. These book names and similarity scores are recorded in a
list. On completion, the list is reviewed, and any duplicates or items
already bought by the user are removed. The remaining items are sorted
by similarity score and the books with the highest scores are presented
to the user as recommendations.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

150 Googling the Web

In summary, the algorithm works as follows:

Take the similarity table and user’s purchasing history as input.

Create an empty list.
Repeat for every item in the purchasing history:
Find the matching row in the similarity table.
Repeat for every column in that row:
If the similarity score is greater than zero,
then add the matching title and score to the list.
Stop repeating at the end of the row.
Stop repeating at the end of the purchasing history.
Remove any duplicates from the list.
Remove any book that the user has already bought.
Sort the list by similarity score.
Output the titles with the highest similarity score.

Let’s say that Charlotte’s Web and The Little Prince have been pair pur-
chased by four different users; Charlotte’s Web and Pinocchio were bought
by one customer; as were The Little Prince and Pinocchio (Table 8.1). Imagine
that Nicola web surfs to the Amazon site. The recommender algorithm
recovers her purchasing history and finds that she has only bought one
book so far: The Little Prince. The algorithm scans along the The Little Prince
row in the similarity table to find two non-zero entries: 4 and 1. Looking
up the corresponding column headers, the algorithm discovers that
Charlotte’s Web and Pinocchio are paired with The Little Prince. Since Charlotte’s
Web has the higher similarity score (4), it is presented to Nicola as the best
recommendation. In other words, Nicola is more likely to buy Charlotte’s
Web than Pinocchio since, in the past, more customers have bought both
The Little Prince and Charlotte’s Web.

Table 8.1 Similarity table for three books. The entries

indicate the number of times that each pair of books has
been purchased by a user.

Charlotte’s The Little Pinocchio

Web Prince
Charlotte’s Web – 4 1
The Little Prince 4 – 1
Pinocchio 1 1 –
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Google Web Search 151

Recommendation accuracy improves significantly with data set size.

The more data in the product similarity table and the more extensive
a user’s history, the better the recommendations get. With more data,
oddities disappear, and major trends emerge. Given enough data, rec-
ommender algorithms are surprisingly accurate. According to a recent
McKinsey report, thirty-five per cent of Amazon’s sales come from
product recommendation.
The year that Amazon launched (1995), the WWW boasted forty-four
million users and twenty-three thousand sites. The following year the
number of users doubled and the number of websites rose by a factor of
ten. The World Wide Web was becoming a frenzy.
The sheer number of webpages began to create problems for users.
Users mindlessly surfed the web, clicking on one link after another. How
could anyone find what they were looking for? Manually curated web-
site directories that sought to catalogue the WWW by topic were swamped.
Finding material online became increasingly slow and frustrating. Users
desperately wanted a website that would magically display exactly the
right link.

Google Web Search

In the spring of 1995, Sergey Brin was asked to show a new guy around
Stanford University. Aged twenty-one, Brin had been a student at Stan-
ford for two years. Born in Moscow, Brin’s parents migrated to the US
when he was six. He graduated with a degree in Computer Science and
Mathematics from the University of Maryland at just nineteen before
winning a scholarship to attend graduate school at Stanford.
The new guy was twenty-two-year-old Larry Page. Page had recently
finished a degree in Computer Science at the University of Michigan.
Hailing from the Midwest, Page was overawed by Stanford. He half
expected to be sent home.
Brin and Page took a near instant dislike to one another. Page: 149
I thought he was pretty obnoxious. He had really strong opinions about
things, and I guess I did, too.
Brin:
We both found each other obnoxious.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

152 Googling the Web

Nonetheless, the two grad students found common ground in a love of

repartee:
We had a kind of bantering thing going.
Page wasn’t sent home on the next bus. He started his PhD. He hung
out with Brin. They duo began to collaborate on projects of mutual
interest.
Everyone in computer science knew that the WWW was hot.
Netscape had just floated on the stock market at a valuation of $3
billion despite having zero profits. The company’s only product was a
web browser, which they gave away for free. Netscape’s Initial Public
Offering (IPO) heralded the beginning of the Wall Street dotcom bubble.
The popular dotcom moniker derived from the URL extension assigned
to commercial web sites. Web start-up after web start-up hit crazy stock
market valuations.
Page’s thoughts turned to the problem of web search. Website directo-
ries clearly weren’t the way to go. Drilling down through alphabetic lists
of categories, then topics, then sub-topics was way too time consuming
for users. Queries made more sense. Just type in what you are looking
for, and the most relevant links should appear. The difficulty was that
matching query terms with the titles of web sites didn’t work at all
well. Most of the links returned were either irrelevant or low quality,
or both. Users were forced to sift through a lot of rubbish before they
found what they were looking for. Page realized that accurately ranking
weblinks based on their importance and relevance to the query was key.
An accurate website ranking system would push useful links to the top
of the list. The question was how to accurately rank websites?
Page was familiar with an existing paper-based system for ranking
academic research papers. Research papers are typically ranked accord-
ing to the number of times that they have been referenced in other
publications. The idea is that the more often a paper is referenced, or
cited, the more important it is. Page realized that hyperlinks are a lot
like references. A hyperlink from one page to another indicates that the
author of the linking page thinks that the linked page is, in some way,
important or relevant. Counting the number of links directed to a page
might be a good way to assess the page’s importance. Working from this
insight, Page developed an algorithm for ranking the importance of web
pages. In a play on his own name, he dubbed the method PageRank.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Google Web Search 153

PageRank does more than simply count citations: it takes into con-
sideration the importance of the pages that link to the page being rated.
This stops webmasters artificially raising the rank of a page by creating
spurious pages linking to it. To have any impact, the linking pages must,
themselves, be important. In effect, PageRank ranks webpages based on
the communal intelligence of website developers.
Every webpage is allocated a PageRank score: the higher the score,
the more important the page. The score for a page is equal to the sum
of the weighted PageRanks of the pages that link to it plus a damping term.
The PageRank of an incoming link is weighted in three ways. First, it
is multiplied by the number of links from the linking page to the page
being scored. Second, it is normalized, meaning that it is divided by the
number of links on the linking page. The rationale is that a hyperlink
from a page that contains many links is worth less than a hyperlink from
a page with a small number of links. Third, the PageRanks are multi-
plied by a damping factor. This damping factor is a constant value (typically
0.85) which models the fact that a user might jump to a random page
rather than follow a link. The damping term compensates for this by
adding one minus the damping factor to the total (usually 0.15).
PageRank can be thought of as the probability that a web surfer who
selects links at random will arrive a particular page. Lots of links to a
page mean that the random surfer is more likely to arrive there. Many
hyperlinks to those linking pages also means that the surfer is more
likely to arrive at the destination page. Thus, its PageRank depends not
only on the number of links to a page but also on the PageRank of
those linking pages. Therefore, the PageRank of a destination web-page
depends on all of the links that funnel into it. This funnelling effect
goes back one, two, three, and more links in the chain. This dependency
makes PageRank tricky to calculate. If every PageRank depends on every
other PageRank, how do we start the calculation?
The algorithm for computing PageRanks is iterative. Initially, the
PageRanks are set equal to the number of direct incoming links to a
page divided by the average number of inward links. The PageRanks
are then recalculated as the total of the weighted incoming PageRanks
plus the damping term. This gives a new set of Page Ranks. These values
are then used to calculate the PageRanks again, and so on. With every
iteration, the calculation brings the PageRanks closer to a stable set of
values. Iteration stops when the PageRanks show no further change.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

154 Googling the Web

Imagine a mini-WWW comprised of just five pages (Figure 8.2). At

first glance, it is difficult to determine which page is most important.
Two pages have three incoming links (B and E). One page has two
inward links (A). The other two pages are less popular, with just one
incoming link (C and D).
Calculation of the PageRank scores begins by creating a table showing
the number of links between pages (Table 8.2). Every row corresponds
to a page, as does every column. The entries in the table record the
number of links from the page in the row to the page in the column.
Thus, reading across a row, we have the number of outgoing links for
the named page. Scanning down a column, we have the number of
incoming links for the named page.

Figure 8.2 Five inter-linked web pages.

Table 8.2 Table showing the number of links between pages. The table also
includes the total number of links. A page may not link to itself.

To A To B To C To D To E Total out
From A - 1 1 0 0 2
From B 0 - 0 0 1 1
From C 0 0 - 1 1 1
From D 1 1 0 - 1 3
From E 1 1 0 0 - 2
Total in 2 3 1 1 3
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Google Web Search 155

The algorithm creates a second table listing the PageRanks for every
page (Table 8.3). To start with, the algorithm populates the table with a
rough estimate of the PageRanks. This is the number of incoming links
for a page divided by the average number of inward links.
The algorithm then recalculates the PageRanks page by page, that
is, column by column in the PageRank table. A page is processed by
calculating the incoming weighted PageRanks for a single column in
the links table. The PageRank of each incoming link is obtained by
looking up the current PageRank for that page (Table 8.3). This value is
multiplied by the number of incoming links from the page. The result
is divided by the total number of outgoing links on the page (Table 8.2).
This value is multiplied by the damping factor (0.85) to obtain the
weighted PageRank. This calculation is performed for all incoming
links. The resulting weighted incoming PageRanks are totalled and
added to the damping term (0.15). This gives the new PageRank for the
page. This value is appended to the PageRank table.
The PageRank calculation is repeated for all of the pages. When all
of the pages have been processed, the new PageRanks are compared to
the previous values. If the change is small, the process has converged
and the results are output. Otherwise, the calculation is repeated (the
complete algorithm is listed in detail in the Appendix).
For example, imagine performing the second iteration in calculating
the PageRank of page A. Its PageRank is the sum of the weighted
PageRanks of the pages with incoming links, that is, pages D and E. The
initial PageRank of D is 0.5. The number of links from D to A is 1. The
number of links from D is 3. Therefore, the weighted PageRank from D
to A is 0.5 × 1 ÷ 3 = 0.167. Similarly, the weighted PageRank from E to
A is 1.5 × 1 ÷ 2 = 0.75. Including the damping term, the new PageRank
for A is 0.167 + 0.75 + 0.15 = 1.067.
The iterative procedure balances the PageRanks so that they reflect
the inter-connection of the webpages. The larger values flow to the

Table 8.3 Table showing the final PageRank values obtained

for the mini-WWW example.

Iteration A B C D E
1 1.00 1.50 0.50 0.50 1.50
5 0.97 1.38 0.56 0.40 1.69
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

156 Googling the Web

more interconnected pages, the smaller to the less well connected.

Eventually, the numbers settle down to a steady-state that balances
the flow.
For the mini-WWW example, the PageRank scores converge after
about five iterations (Table 8.3). In the end, webpage E has the greatest
PageRank. Page E has the same number of incoming links as B but ranks
ahead of it. This is because page B, which links to E, only has a single link.
Whereas, page E, which links to B, has two links. Normalization reduces
the PageRank of page B.
To try out his algorithm, Page, in collaboration with Brin and
Stanford Professor Rajeev Motwani, built a prototype search engine.
They used a web crawler to download a summary of the WWW. A web
crawler explores the WWW much like a human surfer, following link
after link. Along the way, the crawler keeps a snapshot of every web
page that it comes across. Starting with a few manually curated URLs,
a crawler quickly creates a local summary of the WWW. The PageRank
algorithm is applied to the resulting dataset to score the importance
of the pages. On receiving a user query, the team’s search engine—
BackRub—searches the local summary for matching page titles. The
resulting list of pages is sorted by PageRank and displayed to the user.
BackRub was only ever used internally within Stanford. Nonetheless,
its performance was sufficiently promising that Page and Brin decided to
expand the service.
They bought more computers so that a larger portion of the WWW
could be indexed, and a greater number of concurrent queries handled.
On reflection, they decided BackRub was a poor name. They needed
something that intimated the future scale of the WWW. They decided
on ‘Googol’—a googol being a one followed by one hundred zeros.
By mistake, they wrote the word down as ‘Google’. Brin drew up a
multicoloured logo for the new site. By July 1998, the Google search en-
gine had indexed twenty-four million pages. Crucially, Google’s search
results were light years ahead of their competitors’. The new kid on the
block was ready for the big time.
In August, Brin and Page were introduced to Andy Bechtolsheim.
A Stanford alumnus, Bechtolsheim had cofounded two successful tech
start-ups. Brin and Page pitched the Google proposition to Bechtol-
sheim. He liked what he heard. Right there and then he wrote a cheque
for $100,000 made out to Google, Inc. No negotiations, no terms and
conditions, and no valuations. Bechtolsheim just wanted to be part of
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Dotcom Bubble 157

it. In spirit, he felt he was paying back an old favour granted by one of his
early backers. Clutching the cheque, Brin and Page neglected to point
out that Google, Inc., didn’t exist—yet.
The following February, PC Magazine reported that the fledgling
Google search engine: 153
has an uncanny knack for returning extremely relevant results.
One year later, Google, Inc., received $25 million in funding from
venture capitalists Sequoia Capital and Kleiner Perkins. Sequoia Capital
and Kleiner Perkins were, and still are, Silicon Valley royalty. Their
names on the investor sheet were almost worth more than the money.
The following year, Google launched AdWords. AdWords allows
advertisers to bid for their links to be listed on the Google search page
alongside the PageRank results. The promoted links are clearly delin-
eated from the PageRank results so that users can distinguish between
PageRank results and advertisements. AdWords proved to be much
more effective than traditional advertising. This was hardly surprising,
as AdWords was offering products that customers were already actively
searching for. Companies flocked to the new advertising platform.
AdWords turned free web search into a goldmine.
The PageRank algorithm was patented by Stanford University.
Google exclusively licensed the algorithm back from Stanford for use
in its search engine in exchange for 1.8 million shares. In 2005, Stanford
University sold its Google shares for $336 million. That transaction
probably makes PageRank the most valuable algorithm in history.

The Dotcom Bubble

Hot on the heels of the Netscape IPO in 1995, investments poured into
web start-ups. Profits didn’t seem to matter. The only metric for valu-
ation was the number of users that a site could claim. Web investment
became a pyramid scheme. The tech-loaded NASDAQ index of shares
quintupled in value between 1995 and 2000. The index peaked at 5,048
on 10 March 2000. Thereafter, it plummeted. On 4 October 2002, the
NASDAQ index returned to 1,139, shedding more than three quarters of
its value. It was almost back to where is started in 1995. The dotcom bust
hit the tech industry hard. A slew of web start-ups went to the wall. It
would be fifteen long years before the NASDAQ recovered its year 2000
peak.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

158 Googling the Web

The speculative dotcom bubble disguised steady and continuous

growth in Internet usage. Fuelled by an expanding user base, the
surviving web companies grew rapidly.
Linden’s product recommender algorithm was widely emulated. He
left Amazon in 2002. After spells at two start-ups and Google, he now
works at Microsoft. Today, Amazon sells everything ‘from A to Z’. In
2019, Amazon overtook Walmart as the world’s largest retailer. Accord-
ing to Forbes magazine, Jeff Bezos—Amazon founder and CEO—became
the world’s richest person in 2018. His net worth was estimated at
$156 billion.
Google IPOed in 2004 at a valuation of $23 billion. The company was a
mere six years old. In 2016, Google’s parent company—Alphabet—was
valued at almost half a trillion dollars. As of writing (2019), Brin and
Page are among the top ten richest people in the US according to Forbes
magazine. Their estimated personal wealth is $35–40 billion, each.
Tim Berners-Lee was granted a knighthood by the Queen of England
in 2004 in recognition of his achievements. He is holder of the 2016 ACM
Turing Award. The Richest website estimates his net worth at $50 million.
A small fortune, but a far cry from the wealth of the web billionaires.
When the dotcom bubble burst (2002), there were roughly half a
billion active Internet users. The web afforded the world unprecedented
access to information, online shopping, and entertainment. However, a
nineteen-year-old student living in a college dorm was convinced that
what people really wanted was gossip. That insight, combined with a lot
of hard work, would make him a multibillionaire.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

9
Facebook and Friends

With all these data you should be able to draw some just inference.
Sherlock Holmes to Dr Watson
Sir Arthur Conan Doyle
The Sign of Four, 1890 160

Mark Zuckerberg (Figure 9.1) was born in White Plains, New York, in
1984. His father taught him how to program when he was in middle
school. Later, his dad hired a professional programmer to tutor him.
A whiz in high school, Zuckerberg was always destined for the Ivy
League. He enrolled in Harvard University and selected joint honours
in Computer Science and Psychology.
As well as a love of coding, Zuckerberg had an abiding interest in
the behaviour of people. He realized early on that most people are
fascinated by what other people are doing. This obsession lies at the
very heart of everyday gossip, weighty biographies, celebrity culture,
and reality TV. At Harvard, he began to experiment with software
that would facilitate the basic human need to connect and interact
with others.
Zuckerberg set up a website called Facemash. 162 Facemash was mod-
elled on an existing web site by the name of Hot Or Not. 161 Both sites
displayed side-by-side images of two male, or two female, students
and asked the user to select the more attractive of the two. Facemash
collated the votes and displayed ranked lists of the ‘hottest’ students.
Controversially, Facemash used photographs of students downloaded
from Harvard web sites. The site was popular with a certain cohort of
students but upset a lot of others. According to the on-campus newslet-
ter, the site landed Zuckerberg in front of a disciplinary board. 163
Afterwards, Zuckerberg turned to constructing a new website for
social networking. A few such sites already existed, allowing users to share
information about themselves. Most previous sites were aimed at people
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

160 Facebook and Friends

Figure 9.1 Facebook co-founder Mark Zuckerberg, 2012. (By JD Lasica from
Pleasanton, CA, US - Mark Zuckerberg, / Wikimedia Commons / CC BY 2.0, https://commons.
wiki-media.org/w/index.php?curid=72122211. Changed from colour to monochrome.)

seeking dates. Instead, Zuckerberg wanted his social network to help

Harvard students communicate. The idea was that users would enter
personal profiles and post news. This wasn’t the sort of news that would
make the headlines but it was the sort of chit-chat that students loved.
To sidestep university rules, the new site required that users upload
their own data. The new site—thefacebook.com—was launched in
February 2004. Zuckerberg was nineteen.

Facebook’s News Feed

Word about Facebook spread quickly. Four days after launch, it had 450
registered users. Students used the site for all sorts of things—planning
parties, organizing study sessions, and, inevitably, dating. Gradually,
Zuckerberg opened the social network up to students from other US
universities. In June, he was offered $10 million for the site. He wasn’t
remotely interested. As the number of users grew, Zuckerberg took on
investment, began hiring, and dropped out of college.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Facebook’s News Feed 161

In the beginning, the only way to find a fresh post was to check users’
profile pages for updates. Mostly, checking pages was a waste of time—
there was just nothing new to see. Zuckerberg realized that it would
be helpful for users to have a page summarizing the latest posts from
their pals.
Over the next eight months, the Facebook News Feed algorithm was
born. The algorithm proved to be the biggest engineering challenge
that the young company had faced. News Feed wasn’t just a new
feature. It was a re-invention of Facebook.
The idea was that News Feed would produce a unique news page for
every single user. The page would list the posts most relevant to that
particular user. Everyone’s feed would be different – personalized for
them by the system.
Facebook activated News Feed on Tuesday, 5 September 2006. User
reaction was almost unanimous. Everyone hated it. People felt it was
too stalker-esque. In fact, nothing was visible that had not been available
on Facebook previously. However, News Feed made it easier to see what
was going on in everyone else’s lives. It seemed that Zuckerberg had
misjudged users’ emotional reaction to a perceived change in their data
privacy.
Anti-News Feed groups sprang up on Facebook and flourished. Iron-
ically, students were using the very feature that they were objecting to,
to help them protest. For Zuckerberg, this was proof positive that News
Feed had worked. The numbers backed him up. Users were spending
more time on Facebook than ever before. Facebook made its apologies,
added privacy controls, and waited for the fuss to die down.
The engineering challenge at the core of News Feed lay in creating
an algorithm that would select the best news items to display to a user.
The question was: how could a computer algorithm possibly determine
what a human user was most interested in? The details of Facebook’s
News Feed algorithm remain a closely guarded secret. However, some
information was disclosed in 2008.
The original News Feed algorithm was called EdgeRank. The name
seems to have been a nod to Google’s PageRank. Every action on Face-
book is called an edge, be it a user post, a status update, a comment, a
like, a group join, or an item share. An EdgeRank score is calculated for
every user and every edge by multiplying three factors:
EdgeRank = affinity × weight × time decay.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

162 Facebook and Friends

Affinity is a measure of the user’s degree of connection to the edge.

It indicates how close the user is to the person who created the edge.
Friends on Facebook are considered closer than non-friends. The more
mutual friends that two friends have, the greater their affinity. The
number of interactions between users affects their affinity score. For ex-
ample, affinity is boosted if users often comment on each other’s posts.
Affinity wanes over time if users stop interacting with one another.
Weight depends on edge type. Edges that require more effort to create
have a higher weight. A comment, for example, has a greater weight
than a like.
All else being equal, EdgeRank decays as an edge ages. This ensures
that the algorithm prioritizes more recent posts over older items.
Every fifteen minutes, EdgeRanks are re-calculated for every user and
every post. A user’s NewsFeed is prepared by sorting posts in descending
order of their EdgeRank scores. Over time, the relative EdgeRank scores
change. One post may be demoted as it ages. Another’s rank might rise
as it receives a flurry of likes. This dynamism encourages users to keep
checking their feeds in the hope of seeing something new.
News Feed popularized viral messaging. Users cannot broadcast mes-
sages on Facebook. They simply post a message in the hope that it will
propagate to their social network. If an item receives lots of attention,
in the form of likes or comments, then its EdgeRank score is boosted.
As the score increases, the item appears in a wider circle of users’ feeds.
Popular messages can cross between user communities, much like the
spread of a virus.
Facebook parlayed News Feed views to revenue by interspersing user
posts with sponsored advertisements. The company was publicly listed
on the stock exchange in 2012 at a valuation in excess of $104 billion. The
Facebook website and apps now have around 2.4 billion monthly users.
Even allowing for bots (i.e. programs mimicking users) this is a large
proportion of the world’s population. As of 2020, Zuckerberg remains
CEO of Facebook. According to Forbes magazine, Zuckerberg’s net worth
is estimated at around $60 billion.
News Feed offers a form of personalization, tailoring news content to
the individual. While personalization had been available on the web
since before the arrival of Linden’s Amazon recommender, it was
an algorithm developed for Netflix that took the technology to the
next level. That algorithm leveraged machine learning to identify and
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Netflix Prize 163

exploit the latent patterns in large volumes of user data. Machine

learning combined with big data would soon transform business and
science.

The Netflix Prize

Netflix was founded in 1997 by Reed Hastings and Marc Randolph.
Originally natives of the east coast, Hastings and Randolph gravitated
to Silicon Valley. Both navigated the merger and acquisition maelstrom
of the technology industry, becoming serial entrepreneurs. They first
met when Hastings’ company acquired a software start-up for which
Randolph worked. Living close to one another, the colleagues began
to carpool. Over the course of their daily commutes, Hastings and
Randolph hatched a plan for a new business venture.
The business proposition was straightforward—online movie rental.
Subscribers selected movies that they wanted to watch via the com-
pany’s web site. As they became available, the movies were posted out
on disk (DVD). After viewing, customers returned the DVDs by mail.
The service proved popular. Subscribers liked having access to a large
movie library, plus they enjoyed the convenience of DVDs arriving on
their doorstep. The key to success was ensuring that customers actually
enjoyed the movies they received in the post. Following Amazon’s
lead, Netflix added a recommendation engine to their web site. Net-
flix’s recommender, Cinematch, worked well. Nevertheless, by 2006,
the company was looking for something better. Rather than develop
another algorithm in-house, Netflix took the unusual step of launching
a public competition. The company announced a $1 million prize for
the first recommender system that was ten per cent more accurate than
Cinematch.
To facilitate the competition, Netflix released a training dataset con-
taining 100 million movie ratings allotted by nearly half a million
customers to almost 18,000 different movies. Every data point consisted
of the movie’s name, the user’s name, the star rating assigned (1 to 5),
and the date that the rating was logged. The movies’ names and ratings
were real, but the users’ names were anonymized so that individuals
could not be identified.
In addition, a second, qualifying, dataset was released. Its contents
were similar to the training dataset except that Netflix held back the
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

164 Facebook and Friends

star ratings. The qualifying dataset was much smaller, containing just
2.8 million entries.
The goal of the competition was to build a recommender that could
accurately predict the redacted movie ratings in the qualifying dataset.
Netflix would compare the estimates provided by a competitor with the
hidden user assigned ratings. The competitor’s estimates were evaluated
by measuring the prediction error—the average difference between the
predicted and actual ratings squared.
The $1 million prize attracted hobbyists and serious academic re-
searchers alike. As far as the academics were concerned, the dataset was
gold dust. It was very difficult to get hold of real-world datasets of this
size. At the outset, most thought that a ten per cent improvement in
accuracy was going to be trivial. They underestimated the effectiveness
of Cinematch.
Ratings predictions can be made in a large number of ways. The most
effective technique used in the competition was to combine as many
different predictions as possible. In predictor parlance, any information
which can be used as an aid to prediction is a factor that must be taken
into account in the final reckoning.
The simplest factor is the average rating for the movie in the training
dataset. This is the average across all users who have watched that
particular movie.
Another factor that can be considered is the generosity of the user
whose rating is being predicted. A user’s generosity can be calculated
as their average rating minus the average across all users for the same
movies. The resulting generosity modifier can be added to the average
rating for the movie being predicted.
Another factor is the ratings given to the movie by users who gener-
ally score in the same way as the user in question. The training dataset
is searched for these users. The prediction is then the average of their
ratings for the movie.
Yet another factor is the ratings given by the user to similar movies.
Again, the training dataset is inspected. This time, movies typically
allocated similar ratings to the show in question are identified. The
average of the user’s ratings for these films is calculated.
These factors, and any others available, are combined by weighting
and addition. Each factor is multiplied by a numeric value, or weight.
These weights control the relative importance of the factors. A high
weight means that the associated factor is more important in deter-
mining the final prediction. A low weight means that the factor is of
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Netflix Prize 165

less importance. The weighted factors are summed to give the final
prediction.
In summary, the prediction algorithm is then:

Take the training and qualifying datasets as input.

Repeat for every user-movie combination in the qualifying
dataset:
Repeat for every factor:
Predict the user’s movie using that factor.
Stop repeating when all factors have been assessed.
Weight and sum the factor predictions.
Output the final prediction for that user and movie.
Stop repeating when all user-movie combinations have been
predicted.

Imagine that the algorithm is trying to predict Jill’s rating for Toy
Story (Table 9.1). The first factor is simply the average rating for Toy Story
in the training dataset. This gives an estimate of 3.7 stars. The second
factor—Jill’s generosity—is obtained by calculating Jill’s average rating
and subtracting the average rating for the same movies in the training
dataset. Jill’s average rating is 4 while the average for the same movies is
3.1 Therefore, Jill’s generosity bonus is +0.9. Adding this to the global
average gives 4.6 stars. Next, the dataset is searched for users whose
previous ratings are similar to Jill’s. This is clearly Ian and Lucy. They
gave Toy Story 5 stars, so that’s another factor. The fourth factor requires
that movies that Jill has watched which normally attain similar ratings
to Toy Story are found. The obvious ones are Finding Nemo and The Incredibles.
Jill gave these two movies an average rating of 4.5 stars. So, that is
another factor. In conclusion, we have four estimates of Jill’s rating of
Toy Story: 3.7, 4.6, 5, and 4.5 stars.

Table 9.1 Movie ratings dataset.

Toy Finding The Frozen

Story Nemo Incredibles
Ian 5 4 4 2
Jill ? 4 5 3
Ken 1 2 1 4
Lucy 5 4 5 2
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

166 Facebook and Friends

If the last two factors are typically the most reliable, we might well
use the weights: 0.1, 0.1, 0.4, 0.4. Multiplying and summing gives a final
prediction of 4.6 stars. Jill should definitely watch Toy Story!
Although this general approach was common, teams varied in the
specific features they exploited. Sixty features or more were not un-
usual. Teams also experimented with a wide range of similarity met-
rics and ways to combine predictions. In most systems, the details of
the predictions were controlled by numerical values. These numerical
values, or parameters, were then tuned to improve the prediction results.
For example, the weight parameters were tweaked to adjust the relative
importance of factors. Once the factors are identified, accurate rating
prediction depends on finding the best parameter values.
To determine the optimum parameter values, teams turned to ma-
chine learning (see Chapter 5). To begin with, a team sets aside a subset
of the training dataset for validation. Next, the team guesses the param-
eter values. A prediction algorithm is run to obtain predictions for
the ratings in the validation dataset. The prediction error is measured
for these estimates. The parameters are then adjusted slightly in the
hope of reducing the prediction error. These prediction, evaluation, and
parameter adjustment steps are repeated many times. The relationship
between the parameter values and the prediction error is monitored.
Based on the relationship, the parameters are tuned so as to minimize
the error. When no further reduction in error can be obtained, training
is terminated and the parameter values frozen. These final parameter
values are used to predict the missing ratings in the qualifying dataset
and the results submitted to Netflix for adjudication.
Overall, the training algorithm works as follows:

Take the training and validation datasets as input.

Guess the best parameter values.
Repeat the following:
Run the prediction algorithm for all items in the validation
dataset.
Compare the predicted ratings to the actual ratings.
Adjust the parameters to reduce the error.
Stop repeating when no further improvement is obtained.
Output the parameters which give the minimum prediction
error.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

The Netflix Prize 167

The beauty of the machine learning approach is that the computer

can experiment with far more parameter combinations than any hu-
man.
At first, teams progressed rapidly. The top algorithms quickly hit
six, seven, and eight per cent improvements over Cinematch. Then,
accuracy stalled. Teams poured over their validation results to find
out what was going wrong. The stumbling block was an issue that
became known as the Napoleon Dynamite problem. Napoleon Dynamite was
the hardest movie to predict a rating for. Most movie ratings could
be predicted fairly easily because there was something similar in the
dataset. The trouble with Napoleon Dynamite was that there was nothing
quite like it. The film was an oddball independent comedy that had
become a cult hit. People either loved it or hated it. It was the kind of
movie that friends would argue over for hours on end.
Although Napoleon Dynamite was the hardest movie to predict, it
wasn’t alone. There were enough hard-to-predict movie ratings to halt
progress. Some competitors began to wonder if a ten per cent uplift was
even possible.
Two years after the competition started, a handful of competitors
realized that there was a way forward after all. Each team had designed a
custom algorithm. Each algorithm had its own particular strengths and
weaknesses. In many cases, the strengths and weaknesses of competing
algorithms were complementary. Teams realized that accuracy could
be enhanced by combining the estimates from multiple algorithms.
Again, machine learning was employed to determine the best way to
combine the various predictions.
Accuracies began to push upwards once more. The more different
the algorithms, the better the combined ensemble results seemed to be.
Teams rushed to integrate their algorithms with other high performing
solutions. The competition reached its climax in a frenzy of team
mergers and assimilations.
On 21 September 2009, the Netflix Prize was won by BellKor’s
Pragmatic Chaos. The team had achieved a 10.06% improvement on
Cinematch. The group was an amalgamation of KorBell out of AT&T
Research in the US; Big Chaos from Austria; and Pragmatic Theory
from Canada. All in, it was a group of seven people working across
the globe, primarily communicating via email. Of course, the merger
meant that the $1 million prize had to be split seven ways. Still, it was a
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

168 Facebook and Friends

pretty good payday. Pity the second placed team who, after three years,
lost by just ten minutes.
In a plot twist, Netflix decided not to deploy the winning algorithm.
The company had already replaced Cinematch with the winner of an
earlier stage. Netflix reckoned that an 8.43% improvement was good
enough and left it at that.
The competition was a resounding success. In total, 51,051 contes-
tants from 186 countries organized into 41,305 teams had entered.
While the competition was underway, Netflix announced a shift
from disk rental to online streaming of movies across the Internet.
The company soon discontinued its disk rental service entirely. Today,
Netflix is the world’s largest Internet television network, with more
than 137 million subscribers. Marc Randolph left Netflix in 2002. He is
now on the board of several tech companies. Reed Hastings continues
as Netflix CEO. He is on Facebook’s Board of Directors as well. Hastings
is currently at number 504 on the Forbes Billionaires List (2019).
McKinsey recently reported that a staggering seventy-five per cent of
Netflix’s views are based on system recommendations. Netflix estimates
that personalization and movie recommendation services save the com-
pany $1 billion every year through enhanced customer retention.
In 2009, it was starting to seem like big data, coupled with machine
learning, could predict almost anything.

Google Flu Trends

That year, an eye-catching paper in Nature suggested that data ana-
lytics applied to Google web search queries could track the spread
of influenza-like diseases across the US. The concept was intuitively
appealing. People who feel ill often use Google web search to find out
about their symptoms. A spike in user queries related to the flu might
well indicate an outbreak of the disease.
It was a significant announcement. Seasonal flu is a major health
problem. Every year, there are tens of millions of flu cases worldwide,
causing hundreds of thousands of deaths. Furthermore, there is a per-
sistent risk that a new, more virulent, strain of the disease will evolve.
The flu pandemic of 1917–1918 killed somewhere between twenty and
forty million people—more than died in the First World War.
The authors of the paper employed machine learning to deter-
mine the relationship between Google search queries and real-world
flu statistics. The flu statistics were provided by the US Centers for
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Google Flu Trends 169

Disease Control and Prevention (CDC). The CDC monitors outpatient

visits to hospitals and medical centres across the US. The research team
compared regional data on the percentage of flu-related physician visits
with Google queries made in the same area over the same time period.
The algorithm considered fifty million candidate Google queries com-
paring the occurrence pattern of each with the observed flu statistics.
The researchers identified forty-five search queries whose occurrence
patterns closely matched the flu outbreaks in the CDC data. These
search queries were then harnessed to predict the flu.
The group evaluated their algorithm by predicting the percentage
of flu-related physician visits recorded in a completely different time
period. They found that the algorithm consistently, and accurately,
predicted the CDC data one to two weeks ahead of time. The study
was deemed a success. It appeared that the flu could be predicted by
analysing Google searches. To assist medical agencies, Google launched
Google Flu Trends, a website to provide real-time estimates of the
number of flu cases.
The initiative was widely lauded. Big data and machine learning were
delivering real-time medical information at national scale for free. Yet,
some researchers were sceptical. There were concerns about aspects
of the study. The algorithm had been trained using seasonal flu data.
What if the forty-five search queries were associated with the season,
rather than the flu itself? The original paper noted that the query
‘high school basketball’ was well correlated with the CDC flu data.
High school basketball season coincides with winter flu season, but high
school basketball itself doesn’t indicate that anyone really has the flu.
The researchers had excluded basketball queries from the system but
what if other seasonal correlations were hiding in the data?
A non-seasonal outbreak of influenza virus A(H1N1) provided the
first opportunity to test the claims and counter-claims. The epidemic
started in the summer, rather than the winter, and came in two waves.
This time, the algorithm’s predictions did not match the CDC data.
In response, the team modified the training dataset to include seasonal
and non-seasonal flu events. They also allowed less common search
queries. After these modifications, the revised algorithm accurately
predicted both waves of the H1N1 epidemic.
All seemed well until a paper appeared in Nature in February 2013.
The flu season hit early that winter—the earliest since 2003—and was
worse than normal, leading to an unusually large number of fatalities,
particularly among the elderly. Google Flu Trends overestimated the
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

170 Facebook and Friends

peak of the CDC data by more than fifty per cent—a huge error. Soon,
more Google Flu Trends errors were reported. One research group even
demonstrated that Google Flu Trends’ predictions were less accurate
than basing today’s flu predictions on two-week-old CDC data. Amidst
the clamour, Google Flu Trends was discontinued.
What went wrong?
With hindsight, the original work used too little CDC data for train-
ing. There was so little CDC data and so many queries, that some queries
just had to match the data. Many of these matches were random. The
statistics happened to match but the queries were not a consequence of
anyone actually having the flu. In other words, the queries and flu data
were correlated but there was no causal relationship. Second, there was
too little variability in the epidemics captured in the training data. The
epidemics happened at much the same time of year and spread in similar
ways. The algorithm had learned nothing about atypical outbreaks.
Any deviation from the norm and it floundered. Third, media hype
and public concern surrounding the deaths in 2013 probably led to
a disproportionate number of flu queries which, in turn, caused the
algorithm to overshoot.
The bottom line is that machine learning algorithms are only as good
as the training data supplied to them.
The science of nowcasting has moved on significantly since Google
Flu Trends. Real-world conditions are now determined at scale and
low cost by means of networked data-gathering devices and analyt-
ics algorithms. Sentiment analysis applied to Twitter posts is used to
predict box office figures and election outcomes. Toll booth takings are
used to estimate economic activity in real-time. The motion sensors
in smartphones have been monitored to detect earthquakes. Undoubt-
edly, nowcasting of disease epidemics will be revisited in the future with
the aid of more reliable health sensors.
Meanwhile, in 2005, one year before the Netflix Prize was launched,
a group of IBM executives were on the hunt for a fresh computing
spectacular. The 1997 defeat of world Chess champion Garry Kasparov
by IBM’s Deep Blue computer had hit headlines around the world. That
victory was more about computer chip design than novel algorithms.
Nevertheless, the event was a milestone in the annals of the computing.
IBM wanted a sequel for the new millennium. The challenge had to be
well chosen—a seemingly impossible feat that would grab the general
public’s attention. Something truly amazing …
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

10
America’s Favourite Quiz Show

I, for one, welcome our new computer overlords.

Ken Jennings
On Jeopardy, 2011180

IBM’s T. J. Watson Research Center occupies a futuristic building in

Yorktown Heights, New York. The building’s low semi-circular facade
sweeps away from the entrance to vanishing points at the far ends of
the structure. Its three-storey array of black-framed plate glass windows
looks out onto wooded parkland. The Center is famous for a succession
of historic breakthroughs in both electronics and computing. Incon-
gruously, amidst the snow of early 2011, Watson Center was the venue
for a TV game show.
Jeopardy is an institution in the US. The show has aired almost con-
tinuously since 1964. In every episode, three contestants compete on
the buzzer to win cash prizes. The show’s unique feature is that the
quizmaster gives the ‘answers’ and the players provide the ‘questions’.
In truth, the presenter’s ‘answers’ are cryptic clues. The players’ replies
are merely phrased as questions. For example, the clue: 182
One legend says this was given by the Lady of the Lake and thrown back
in the lake on King Arthur’s death.
should elicit the response:
What is Excalibur?
In 2011, IBM had skin in the game. The company had developed
a Jeopardy-playing computer by the named of Watson. Watson was to
compete against the two best Jeopardy players in history—Ken Jennings
and Brad Rutter. The prize was a cool $1 million.
Ken Jennings held the record for the longest unbeaten winning
streak—74 games on the trot. Along the way, he had accumulated
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

172 America’s Favourite Quiz Show

$2.5 million in prize money. Jennings, now 36, had been a computer
programmer before his Jeopardy success.
Brad Rutter had amassed the greatest winnings in Jeopardy history—
$3.25 million. Rutter was four years Jennings’ junior and worked in
a record store before his first appearance on the show.
The seeds of the IBM Jeopardy Challenge were sown six years previ-
ously. At the time, IBM management were on the lookout for a spec-
tacular computing event. They wanted something that would capture
the public imagination and demonstrate the capabilities of IBM’s latest
machines.
With this in the back of his mind, Director of IBM Research Paul Horn
happened to be on a team night out at a local restaurant. Mid-meal
the other diners left their tables en mass and congregated in the bar. He
turned to his colleagues and asked, ‘What’s going on?’. Horn was told
that everyone else was watching Jeopardy on TV. Ken Jennings was on his
record-breaking winning streak. Half the country wanted to see if he
could keep it going. Horn paused and wondered if a computer could
play Jeopardy.
Back at base, Horn pitched the idea to his team. They didn’t like
it. Most of the researchers reckoned that a computer wouldn’t stand
a chance at Jeopardy. The topics were too wide ranging. The questions
were too cryptic. There were puns, jokes, and double meanings—all
stuff that computers weren’t good at processing. Regardless, a handful
of staff decided to give it a go.
One of the volunteers, Dave Ferrucci, later became Principal Inves-
tigator on the project. A graduate of Rensselaer Polytechnic Institute
in New York, Ferrucci had joined IBM Research straight out of college
after gaining a PhD in Computer Science. His specialism was knowledge
representation and reasoning. He was going to need that expertise—
this was the toughest natural language processing and reasoning chal-
lenge around.
IBM’s first prototype was based on the team’s latest research. The
machine played about as well as a five-year-old child. The Challenge
wasn’t going to be easy. Twenty-five IBM Research scientists were to
spend the next four years building Watson.
By 2009, IBM were confident enough to call the producers of Jeopardy
and suggest that they put Watson to the test. The show executives
arranged a trial against two human players. Watson didn’t perform
at all well. Its responses were erratic—some were correct, others
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

America’s Favourite Quiz Show 173

ridiculous. Watson quickly became the butt of the quizmaster’s

jokes. The computer was the fool in the room. The machine wasn’t
ready.
IBM tried again one year later. This time, Watson gained the approval
of the TV show producers.
Game on.
The contest was recorded at Watson Center and subsequently broad-
cast over three consecutive days (14–16 February 2011).
The host is Jeopardy regular, Alex Trebek. Bedecked in a grey suit,
pink shirt, red tie, and wire frame glasses, Trebek is centre stage—the
epitome of calm sophistication. His grey hair is neatly cut, his brown
eyes sharp, and alert. His smooth vocal tones command attention.
A giant screen to the left of the garish purple and blue set displays
the game board. To the right, the players stand behind individual podia
emblazoned with their names and prize money totals.
Jennings and Rutter flank a computer monitor displaying an an-
imated graphic. The graphic—a blue cartoon of the world crowned
with exclamation marks—is Watson’s visible presence—the computer’s
avatar. The machine’s robotic thumb rests ominously on the buzzer. Jen-
nings sports a yellow tie, lilac shirt, and dark jacket. His ginger-brown
hair is side parted. Rutter’s shirt is open necked beneath a black jacket
adorned with a stylish pocket square. Rutter’s hair is dark brown. His
facial hair is somewhere between designer stubble and a full bread. The
auditorium is packed with IBM top brass, researchers, and engineers.
The highly partisan crowd is charged, noisy, and enthusiastic. This is a
home game for Watson.
Rutter picks a category. Trebek reads the first clue of the day. Simul-
taneously, an equivalent text file is fed to Watson. The clue is: 185
Four-letter word for a vantage point or a belief.
Rutter is first on the buzzer:
What is a view?
Correct. $200 to Rutter. Trebek: 185
Four letter word for the iron fitting on the hoof of a horse or a card dealing
box in a casino.
Watson is first this time:
What is a shoe?
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

174 America’s Favourite Quiz Show

Correct. $400 to Watson. A television camera picks out Ferrucci’s smil-

ing face in the audience.
The clues range from the Beatles to the Olympics. At the end of the
first game, the scores are close. Jennings lags on $2,000, and Watson and
Rutter are tied on $5,000 apiece.
In game two, Watson starts well but some of its replies are distinctly
odd. In response to the clue:
US Cities: Its largest airport is named for a World War II hero, its second
largest for a World War II battle.
Watson replies:
What is Toronto?
The correct answer is Chicago. Toronto isn’t even in the US.
Nonetheless, Watson wins the game. The final scores for game two
are: Watson $35,734, Rutter $10,400, and Jennings a paltry $4,800.
All is still to play for in game three. Early on, Jennings and Watson
are neck and neck, out in front of Rutter. Watson gets a clue right and
hits a Daily Double. Ferrucci punches the air in delight. Jennings visibly
implodes. Later, he says: 185
That’s the moment when I knew it was over.
The final accumulated scores are Watson $77,147, Jennings $24,000,
and Rutter $21,600. IBM Watson wins. In the circumstances, IBM do-
nates the $1 million first prize to charity.
After the match, Rutter opines: 185
I would have thought that technology like this was years away.
For the first time, it seems that conversational artificial intelligence
might be within grasp. Jennings offers: 185
I think we saw something important today.
How did Watson achieve the seemingly impossible?
Certainly, Watson’s processing power and prodigious memory were
part of the computer’s success.
Watson’s hardware was bang up-to-date. The machine consisted of a
network of one hundred IBM Power 750 servers with a total of fifteen
TeraBytes of memory, and 2,880 processor cores. At full tilt, the device
could perform eighty trillion calculations per second.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Watson’s Secret Recipe 175

Watson was in possession of a vast trove of data. The rules required

that the machine be disconnected from the Internet while the match
was in progress. During development, the team had downloaded
one million books to Watson. All sorts of important documents were
crammed into its memory—textbooks, encyclopaedias, religious texts,
plays, novels, and movie scripts.
Despite this, the real secret of Watson’s success lay in its algorithms.

Watson’s Secret Recipe

Watson’s software is an amalgam of hundreds of cooperating algo-
rithms. To begin with, a parser algorithm breaks the clue up into its
constituent grammatical components. The parser determines what
part of speech every word in the clue belongs to. This is done by looking
up the words in a dictionary.
Given the clue: 188
Poets & Poetry: He was a bank clerk in the Yukon before he published Songs
of a Sourdough in 1907.
Watson spots that ‘he’ is a pronoun, ‘was’ is a verb, and ‘bank clerk’ is a
compound noun.
Based on the identified sentence structure, the parser extracts the
relationships between the words. For example, Watson detects a ‘was’
relationship between ‘he’ and ‘bank clerk’, and a ‘publish’ linkage be-
tween ‘he’ and ‘Songs of a Sourdough’. In addition, it spots an ‘in’ relation-
ship between ‘he’ and ‘Yukon’.
After finding the explicit relationships between words, Watson hunts
for implicit links. The original terms are looked up in a thesaurus to
find synonyms. This provides deeper insights into the meaning of the
clue. For example, the ‘publish’ relationship implies an ‘author of’ link.
Once the relationships have been extracted, the elements of the clue
are identified. This is done using a set of if-then-else rules applied to
the parser output. Three main elements are identified: the clue focus,
the answer type, and the question classification. The focus of the clue is the
person, event, or thing that the clue is guiding the contestants towards.
The answer type is the nature of the focus. The question classification is
the category that the clue belongs to. Possible categories include factoid,
definition, multiple-choice, puzzle, and abbreviation. In the example,
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

176 America’s Favourite Quiz Show

the focus is ‘he’—an individual male; the answer type is ‘clerk’ and
‘writer’ (implicit); plus the question classification is ‘factoid’—a short
factual piece of information.
Once clue analysis is complete, Watson searches for answers in its
database. Watson launches a number of searches. These searches access
the structured and unstructured data held in Watson’s memory banks. Struc-
tured data is the name for information held in well-organized tables.
Structured tabular data is great for factoid lookup. For example, Watson
can look up Songs of a Sourdough in a table containing the titles and writers
of well-known songs. However, given the obscurity of the poem, this
hunt is likely to be fruitless.
Unstructured data is the term for information which is not formally
organized. Unstructured data includes information held in textual doc-
uments, such as newspapers or books. Plenty of knowledge is contained
therein but it is difficult for a computer to interpret. Retrieving useful
information from unstructured data turned out to be one of the biggest
problems in building Watson. In the end, the team found some surpris-
ingly effective tricks.
One technique involves searching for an encyclopaedia article men-
tioning all of the words in the clue. Often, the title of the article is the
sought-after answer. For example, searching Wikipedia for the words
‘bank clerk Yukon Songs of a Sourdough 1907’ returns an article entitled
‘Robert W. Service’. This is the correct answer.
Another option is to search for a Wikipedia entry whose title is the
focus of the clue. The algorithm then hunts for the desired information
in the body of the selected article. For example, Watson handles the
clue ‘Aleksander Kwasniewski became the president of this country
in 1995’ by looking up an article entitled ‘Aleksander Kwasniewski’
in Wikipedia. The computer scans the article for the most frequently
occurring country name.
Watson launches a battery of such searches in the hope that one will
yield accurate results.
The resulting candidate answers are assessed by calculating how well
the answers meet the requirements of the clue. Every aspect of the
answers and the clue are compared and scored. The answer with the
highest score is selected as the best solution. The score is compared with
a fixed threshold. If the score exceeds the threshold, Watson rephrases
the solution as a question and presses the buzzer. If called on, Watson
offers the question to the quizmaster.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Watson’s Secret Recipe 177

Watson’s roots lie in the expert systems and case-based reasoning technolo-
gies of the 1970s and 1980s.
Expert systems use hand written if-then-else rules to transform tex-
tual inputs into outputs. The first popular expert system, MYCIN, was
developed by Edward Feigenbaum’s team at Stanford University. It was
designed to assist physicians in determining whether an infection is
bacterial or viral. Bacterial infections can be treated with antibiotics,
whereas viral infections are unresponsive to medication. Doctors com-
monly over-prescribe antibiotics, mistakenly recommending them for
viral infections. MYCIN assists the prescribing physician by asking a
series of questions. These probe the patient’s symptoms and the out-
comes of diagnostic tests. The sequence of questions is determined by
lists of hand-crafted rules embedded in the MYCIN software. MYCIN’s
final diagnosis—bacterial or viral—is based on a set of rules defined by
medical experts.
Case-based reasoning (CBR) systems allow for more flexible decision-
making than expert systems. The first working CBR system is widely
regarded to have been CYRUS, developed by Janet Kolodner at Yale
University. CYRUS is a natural language information retrieval system.
The system holds the biographies and diaries of US Secretaries of State
Cyrus Vance and Edmund Muskie. By referral to these information
sources, CYRUS enters into a dialog with the user, answering questions
about the two subjects. For example: 190
Question: Who is Cyrus Vance?
Answer: Secretary of State of the United States.
Q: Does he have any kids?
A: Yes, Five
Q: Where is he now?
A: In Israel
CYRUS generates candidate answers by matching the query with
passages from the documents. All matches found are scored according
to similarity. The candidate answer with the greatest score is phrased
appropriately and returned to the user.
The main drawback with expert systems is that every rule and con-
sideration must be manually programmed into the system. Case-based
reasoning, on the other hand, requires that every potential linguistic
nuance in the questions and source material is dealt with by program.
Due to the complexity of natural language, the mapping between a
question and valid answer is complex and convoluted. Algorithms must
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

178 America’s Favourite Quiz Show

deal with a bewildering variety of sentence structures. Puns, jokes,

and double meanings make the problem even worse. Every subtlety
of language presented in the input text, and the reference documents,
makes the CBR algorithm more complex.
Watson’s success in playing Jeopardy hinged on the clues. Jeopardy em-
ploys a large, but limited, number of question types. With a Herculean
effort, the team managed to write algorithms to deal with Jeopardy’s most
common clue styles. Had the Jeopardy producers suddenly changed the
format of the clues, Watson would have struggled. In contrast, Watson’s
human opponents would, mostly likely, have adapted. Watson was
programmed to deal with Jeopardy clues and nothing else. Despite the
appearance of genius, Watson understands nothing. It simply shuffles
words according to predefined rules. It was just that Watson had more
rules and data than any previous natural language processing system.
Yet, IBM didn’t build Watson just for playing Jeopardy: 185
[Watson is] about doing research in deep analytics and in natural language
understanding. This is about taking the technology and applying it to
solve problems that people really care about.
After the Challenge was over, IBM set up a business unit to com-
mercially exploit the technologies developed in building Watson. That
business unit is now focused on healthcare applications, especially auto-
mated diagnosis of medical conditions. Retargeting of Watson’s highly
customized algorithms has proven tricky, though. Even IBM executives
have admitted that progress has been slower than expected. 193
With hindsight, the version of Watson that competed in the Jeopardy
Challenge was the pinnacle of case-based reasoning. Much more pow-
erful AI technologies were waiting in the wings. Hints of the coming
revolution were embedded in Watson. Here and there, a handful of
tiny artificial neural networks augmented Watson’s decision-making
capabilities. These networks were portents of the future. The dinosaurs
of AI—expert systems and CBR—were about to be swept away by a
tsunami.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

11
Mimicking the Brain

The machinery of recall is thus the same as the machinery of

association, and the machinery of association, as we know, is
nothing but the elementary law of habit in the nerve-centres.
William James
The Principles of Psychology, 1890 194

Humans have an innate ability to recognize patterns. In just a few for-

mative years, children learn to recognize faces, objects, voices, smells,
textures, and the spoken word. Throughout the twentieth century,
researchers failed miserably in their attempts to design algorithms
that could match human prowess in pattern recognition. Why were
computers so good at arithmetic, yet so poor at pattern recognition?
To better understand the conundrum, imagine developing a system
to recognize cats in photographs.
The first step is to convert an image to an array of numbers that
a computer can process. The lens of a digital camera focuses light onto a
grid of electronic sensors. Each sensor converts the light level to a num-
ber in the range 0 to 1. In a greyscale image, 0 indicates black, 1 white, and
the values in-between shades of grey. Every number corresponds to a
dot, or pixel (picture element), in the image. The array of numbers is
a digital approximation of the image. This much is straightforward. The
challenge in pattern recognition is not in creating a digital image, but
in writing an algorithm capable of making sense of the numbers.
The difficulty arises from the variability of real-world images. First,
there are lots of breeds of cat. The cat could be fat or thin, big or small,
furry or hairless, grey, brown, white, or black, tailed or not. Second,
the cat could be in any one of a great many poses—it might be lying
down, sitting, standing, walking, or jumping. It could be looking at the
camera, to the left, to the right, or have its back to the lens. Third,
the photograph could have been taken under a range of conditions.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

180 Mimicking the Brain

It might be daylight or night, flash photography, a close-up, or a long-

shot. Writing an algorithm to cope with each and every circumstance
is extremely difficult. Every possibility requires a new rule which has to
interact with all of the old rules. Pretty soon, the rules begin to conflict.
Eventually, algorithm development grinds to a halt.
Rather than write millions of rules, a handful of computer scientists
pursued an alternative line of enquiry. Their contention was simple.
If the best pattern recognition engine in the world is the human brain,
why not just replicate it?

Brain Cells
By the onset of the twentieth century, neuroscientists had established
a basic understanding of the human nervous system. The work was
spearheaded by Spanish neuroscientist Santiago Ramón y Cajal, who
received the Nobel Prize for Medicine in 1906.
The human brain is composed of around 100 billion cells, or neurons.
A single neuron is made up of three structures: a central body, a set of
input fibres called dendrites, and a number of output fibres called axons.
When inspected under a microscope, the long thin wispy dendrites and
axons stretch away from the bulbous central body, branching as they
go. Every axon (output) is connected to the dendrite (input) of another
neuron via a tiny gap, called a synapse. Neurons in the brain are massively
interconnected. A single neuron can be connected to as many as 1,500
other neurons.
The brain operates by means of electrochemical pulses sent from one
neuron to another. When a neuron fires, it sends a pulse from its central
body to all of its axons. This pulse is transferred to the dendrites of
the connected neurons. The pulse serves to either excite or inhibit the
receiving neurons. Pulses received on certain dendrites cause excitation,
and pulses received on others cause inhibition. If a cell receives sufficient
excitation from one or more other neurons, it will fire. The firing of one
cell can lead to a cascade of firing neurons. Conversely, pulses arriving
at inhibitory neural inputs reduce the level of excitation, making it less
likely that a neuron will fire. The level of excitation, or inhibition, is
influenced by the frequency of the incoming pulses and the sensitivity
of the receiving dendrite.
Canadian neuropsychologist Donald Hebb discovered that when a
neuron persistently fires, a change takes place in the receiving dendrites.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Brain Cells 181

Figure 11.1 Walter Pitts. (Courtesy Science Photo Library.)

They become more sensitive to the firing neuron. As a consequence,

the receiving neuron becomes more likely to fire in response. Hebb’s
discovery revealed a learning effect in biological neural networks whereby
past experience determines future action. The finding offered a crucial
link between the workings of individual neurons and the higher-level
learning capability of the brain.
In 1943, two American researchers suggested that the workings of
neurons could be modelled by means of mathematics. The proposal was
as radical as it was intriguing.
Walter Pitts (Figure 11.1) was a child genius. Coming from a disad-
vantaged background in Detroit, Pitts taught himself mathematics,
Greek, and Latin from the books in the public library. Browsing the
aisles one day, he spotted Principia Mathematica. Far from an easy read, the
tome establishes the logical foundations of mathematics. In reading the
book, Pitts espied a number of flaws. He wrote a letter to Bertrand
Russell, one of the authors, pointing out the mistakes. Delighted,
Russell replied, inviting Pitts to study with him at Cambridge
University. Unfortunately, Pitts couldn’t go. He was twelve years old.
At fifteen, Pitts heard that Russell was scheduled to speak at
the University of Chicago. Pitts ran away from home, never to return.
He snuck into Russell’s lectures and subsequently took a menial job
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

182 Mimicking the Brain

at the college. Homeless, Pitts happened to befriend Jerome Lettvin,

a young medical student. Lettvin saw something special in Pitts and
introduced him to Warren McCulloch.
McCulloch, a Professor at the University of Illinois, was Pitts’ senior
by twenty-four years. In stark contrast to Pitts’ childhood, McCulloch
was raised in a well-to-do, professional, east coast family. He studied
psychology before obtaining a degree in Neurophysiology. The unlikely
pair, McCulloch and Pitts, entered into deep conversations about Mc-
Culloch’s research. At the time, McCulloch was attempting to rep-
resent the function of neurons by means of logical operations. Pitts
grasped McCulloch’s intent and suggested an alternative, mathemat-
ical approach to the problem. Seeing the young man’s potential and
predicament, McCulloch invited both Pitts and Lettvin to live with him
and his wife.
Ensconced in this new abode, Pitts worked late into the night with
McCulloch. They developed the idea that the state of a neuron could
be represented by numbers. Furthermore, they postulated that the
firing patterns of interconnected neurons could be simulated by means
of equations. They developed mathematical models demonstrating
that networks of neurons could perform logical operations. They even
showed that such networks could assume some of the functions of a
Turing machine (see Chapter 3).
Aided by McCulloch, Pitts procured a graduate position at MIT,
despite the fact that he had not finished high school. Professor of
Mathematics Norbert Wiener, one of his mentors at MIT, guided Pitts
towards cybernetics, the study of self-regulating systems. The subject
encompasses all kinds of self-regulating systems, from biology to ma-
chines. Perhaps the most common examples are today’s thermostati-
cally controlled heating systems. Pitts, Weiner, McCulloch, and Lettvin
formed a loose affiliation to further the field. They reached out to like-
minded researchers such as John von Neumann. Later, Lettvin said of
the cyberneticians: 199
[Pitts] was in no uncertain terms the genius of our group. He was abso-
lutely incomparable in the scholarship of chemistry, physics, of every-
thing you could talk about—history, botany, etc. When you asked him a
question, you would get back a whole textbook. To him, the world was
connected in a very complex and wonderful fashion.
In 1952, McCulloch was invited by Wiener to lead a neuroscience
project at MIT. He jumped at the chance to work with Pitts once again
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Artificial Neural Networks 183

and promptly moved to Boston. Lamentably, Pitts began to struggle

with bouts of melancholy coupled with alcohol addiction. 200 Abruptly
and without explanation, Weiner cut off all contact with Pitts, McCul-
loch, and Lettvin. Floundering, Pitts slipped into depression. He died
from an alcohol-related illness in 1969 at just 46 years of age. McCul-
loch passed away four months later from a heart condition, aged 70.
Their legacy was the mathematical foundation for the artificial neural
network.

Artificial Neural Networks

The world’s first artificial neural network (ANN) was built by Belmont
Farley and Wesley Clark at MIT in 1954. The pair constructed a simple
simulation of firing neurons in a computer program. They used num-
bers to represent the state of the neurons and track the sensitivities of
the inputs and outputs. The network was programmed to recognize
images of simple binary numbers (sequences of ones and zeros). While
Farley and Clark were first, it was Frank Rosenblatt (Figure 11.2) who
popularized the ANN concept.
Rosenblatt was from New Rochelle, New York (born 1928). He stud-
ied at Cornell University, obtaining a PhD degree in Psychology before
taking up a faculty position at the university. To begin with, he simu-
lated an ANN on an IBM computer (1957). Later, to increase the speed

Figure 11.2 Frank Rosenblatt, inventor of the Mark 1 Perceptron, 1950.

(Deceased Alumni files, #41-2-877. Division of Rare and Manuscript Collections, Cornell
University Library.)
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

184 Mimicking the Brain

of processing, he built an ANN in the form of an electronic device:

the Mark I Perceptron (1958). Rosenblatt’s Perceptron was designed to
recognize simple shapes in images. The images (just 20x20 pixels) were
fed from a black and white camera into the Perceptron. The Perceptron
simulated the behaviour of a small network of neurons. The network
output comprised of a set of neuron outputs, each corresponding to
one of the shapes to be recognized. The highest output value indicated
which shape had been spotted in the image.
Although the Perceptron was very limited, Rosenblatt was a persua-
sive communicator. After one of his press conferences, the New York Times
wrote: 205
The Navy revealed the embryo of an electronic computer today that
it expects will be able to walk, talk, see, write, reproduce itself and be
conscious of its existence. Later Perceptrons will be able to recognize
people and call out their names and instantly translate speech in one
language to speech and writing in another language, it was predicted.
The gap between Rosenblatt’s predictions and reality was immense.
Nevertheless, his demonstrations, papers, and book were highly
influential in spreading the concept of ANNs to other research
institutes.
A Perceptron is a classifier—it determines which class, or category, a
given input belongs to. In Rosenberg’s case, the input to the Perceptron
was an image of a shape. The goal of the Perceptron was to determine
whether the shape was a circle, a triangle, or a square. These shapes were
the recognized classes.
A Perceptron consists of one or more layers of artificial neurons. Each
neuron has a number of inputs and a single output (Figure 11.3). The
strength of the signal on an input or output is represented by a number.
The output from a neuron is calculated based on its inputs. Each input is
multiplied by a weight. The weight models the sensitivity of the neuron
to signals on that particular input. The weighted inputs are summed
together with a bias value to give the neuron’s excitation. The excitation is
applied to an activation function. In a Perceptron, the activation function
is a simple threshold operation. If the excitation is greater than the
threshold (a fixed numeric value), then the output is 1. If it is below the
threshold, the output is 0. The threshold value is fixed for the entire
network. The 0 or 1 arising from the activation function is the final
output of the neuron.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Artificial Neural Networks 185

Figure 11.3 Artificial neuron.

The weights and biases of the neurons are collectively called the
parameters of the network. Their values determine the conditions under
which the neurons fire, that is, output a 1. Parameters can be positive or
negative numbers. An input of 1 applied to a positive weight increases
the excitation of the neuron, making it more likely to fire. Conversely,
an input of one multiplied by a negative weight decreases the neu-
ron’s excitation, making it less likely to fire. The bias determines how
much input excitation is needed before the neuron exceeds the fixed
threshold and fires. During normal network operation, the parameters
are fixed.
Perceptrons take a number of inputs (Figure 11.4). For example,
Rosenblatt’s 20x20 pixel image is input to the Perceptron on 400 con-
nections. Each connection has a 0 or 1 value depending whether the
associated pixel is black or white. The input connections are fed into
the first layer of the network: the input layer. Neurons in the first layer
only have a single input connection. Thereafter, the outputs from one
layer feed into the inputs of the next layer. In a fully connected network
all outputs from one layer are input to every neuron in the following
layer. The output of the input layer is connected to the first hidden layer.
Hidden layers are those not directly connected to either the network
inputs or outputs. In a simple network, there might be only one hidden
layer. After the hidden layers comes the output layer. The outputs from
the neurons in this layer are the final network outputs. Each output
neuron corresponds to a class. Ideally, only the neuron associated with
the recognized class should fire (i.e. output a 1).
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

186 Mimicking the Brain

Figure 11.4 Perceptron with three inputs, two fully connected layers, and two
output classes.

In summary, an ANN simulation proceeds as follows:

Take the network input values.

Repeat for every layer:
Repeat for every neuron in the layer:
Set the total excitation equal to the bias.
Repeat for every input to the neuron:
Multiply the input value by the input weight.
Add to the total excitation.
Stop repeating when all the inputs have been processed.
If the excitation is greater than the threshold,
then set the neuron output to 1,
else set the neuron output to 0.
Stop repeating when the layer has been processed.
Stop repeating when all layers have been processed.
Output the class name associated with the greatest network
output.

The overall operation of an ANN can be visualized by imagining

the neurons lighting up when they fire. An input is applied to the
network. This drives 1s and 0s into the input layer. Some of the neurons
fire, sending 1s into the hidden layer. This excites some of the hidden
neurons and they fire as a result. This stimulates one of the neurons
in the output layer. This single output neuron fires, indicating which
pattern was observed at the network input.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Artificial Neural Networks 187

Every neuron makes a tiny decision about the input. Working in

cooperation, a large number of interconnected neurons can make com-
plex decisions about the nature of the input. The capacity to recog-
nize complex patterns arises from the orchestration of all these tiny
decisions.
There are two challenges in ANN design. The first is selecting a
suitable topology. This is the arrangement of neurons in the network. The
topology determines the number of layers, the quantity of neurons in
each, and their interconnection. The topology affects the complexity of
the task that the network can handle. Generally speaking, additional
neurons and layers are needed for more complex pattern recognition
tasks. The second challenge is determining the values for the param-
eters of the network. The parameters control the behaviour of the
network. If the network is to correctly classify the inputs, the parameter
values must be just right.
Rosenblatt chose his network topology based on experience coupled
with trial-and-error, something that hasn’t really changed to this day.
For a given topology, Rosenblatt employed a training procedure
to find effective parameter values. Training started with random pa-
rameters. Rosenblatt then fed a few example inputs into the network
and checked to see if the output was correct. If it wasn’t correct, he
tweaked the parameters until the Perceptron gave the correct answer.
He repeated this process for a set of input examples. Once training was
finished, Rosenblatt assessed the accuracy of his Perceptron by testing
it with previously unseen inputs. Rosenblatt referred to his training
method as back-propagating error correction. More casually, it might be said
that he twiddled the knobs on his Perceptron until it worked. The
method was laborious but gave reasonably good results in recognizing
a small number of simple shapes.
Perceptrons ran into criticism and controversy in the late 1960s. Mar-
vin Minsky and Seymour Papert published a book entitled Perceptrons that
poured cold water on the whole idea. Their book carried weight in the
research community. Both authors were well regarded MIT Professors
and Directors of MIT’s Artificial Intelligence Laboratory. Papert, a South
African, was a mathematician by training. Before arriving at MIT, he
had accumulated an impressive CV, including sojourns at the Univer-
sity of Cambridge, the University of Paris, the University of Geneva,
and the NPL. Minsky, from New York City, had originally been a neural
network believer. His PhD thesis at Princeton was on that very topic. He
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

188 Mimicking the Brain

even designed an electronic device with learning capability inspired by

brain synapses: the SNARC (1952). Thereafter, Minsky became a convert
to the school of symbolic logic (see Chapter 5). For him, neural networks
were out, and logical reasoning was in.
Their book provided a mathematical analysis of the properties of
Perceptrons. It described the strengths of Perceptrons but highlighted
two important limitations. First, they pointed out that single-layer
Perceptrons cannot perform certain elementary logical operations. Sec-
ond, they argued that Rosenblatt’s training method would not work for
multilayer Perceptrons. The book was scathing: 206
Perceptrons have been widely publicized as ‘pattern recognition’ or
‘learning’ machines and as such have been discussed in a large number of
books, journal articles, and voluminous ‘reports’. Most of this writing […]
is without scientific value. Many of the theorems show that Perceptrons
cannot recognize certain kinds of patterns.
The proponents of Perceptrons bristled. Minsky and Papert’s two
central criticisms were valid but misleading, they claimed. True, single-
layer Perceptrons cannot learn certain primitive logic functions, but
multi-layer Perceptrons can. True, Rosenblatt’s training method does
not work for multi-layer Perceptrons but that does not mean that a
suitable training procedure cannot be found. In a review of the book,
H.D. Block retorted: 208
A Perceptron, as defined by Minsky and Papert, is slightly more general
than what Rosenblatt called a simple Perceptron.
On the other hand, the simple Perceptron […] is not at all what a
Perceptron enthusiast would consider a typical Perceptron. He would be
more interested in Perceptrons with several layers, feedback, and cross
coupling. In summary then, Minsky and Papert use the word Perceptron
to denote a restricted subset of the general class of Perceptrons.
Some observers concluded that Minsky and Papert were intentionally
trying to kill the Perceptron. Whatever their rationale, the book con-
tributed to a decline in Perceptron research. Amidst growing disen-
chantment with artificial intelligence, the field as a whole was starved
of funds and the first AI Winter set in.
Tragically, Rosenblatt died in 1971 on his 43rd birthday in a sailing
accident in Chesapeake Bay. Minsky and Papert dedicated the second
edition of Perceptrons to his memory.
In the years that followed, Minsky remained a prolific and celebrated
AI researcher and writer. He won the ACM Turing Award in 1969 for
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

How to Train a Brain 189

his contributions to advancement of the field. Minsky and Papert passed

away in 2016, both aged 88.

How to Train a Brain

After 1970, only a handful of computer scientists persevered with ANNs.
The big problem was finding a way to train multilayer networks. In
the end, the problem was solved four times over by four independent
research teams. Communication was so poor that no one knew what
anyone else was doing. It would be 1985 before word got out that
multilayer ANNs could be trained by means of an algorithm.
At Harvard, Paul Werbos solved the problem as part of his PhD. His
algorithm—back-propagation or backprop—had been around for a while
but had not previously been applied to ANN training. Werbos told
Minsky—by then the doyen of AI—about his solution: 214
In the early 1970s, I did in fact visit Minsky at MIT. I proposed that we do
a joint paper showing that [multilayer Perceptrons] can in fact overcome
the earlier problems. But Minsky was not interested.
More than a decade later (1985), David Parker described backprop in
an MIT technical report. That same year, French student Yann LeCun
published a paper describing an equivalent method at a conference
in Paris. The year after, a letter describing backprop appeared in the
prestigious journal Nature. The authors were David Rumelhart and
Ronald Williams (University of California, San Diego), and Geoffrey
Hinton (Carnegie Mellon University). Unaware of the previous publi-
cations, the group had been working on the idea for a few years. The
letter clearly set out the backprop algorithm and its application to
ANNs. With the imprimatur of Nature behind it, backprop was finally
established as the algorithm for training ANNs.
Backprop requires a minor change to the activation function of
artificial neurons. The threshold operation must be replaced by a smoother
function. The new function ensures that the neuron output rises grad-
ually from 0 to 1 as the excitation increases. Gone is the sudden thresh-
olded transition from 0 to 1 used in the Perceptron. The smooth 0 to 1
transition allows the network parameters to be progressively adjusted
during backprop.
This smoother activation function also means that the network
outputs have a range of possible values between 0 and 1. Thus, the final
decision is no longer the output connection with a 1 value. Instead, it is
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

190 Mimicking the Brain

the output connection with the largest value. The change has the side
effect of improving robustness when the input is on the cusp of two
classes.
Normal operation of an ANN is called forward-propagation (or inference).
The ANN accepts an input, processes the values neuron-by-neuron,
layer-by-layer, and produces an output. During forward-propagation,
the parameters are fixed.
Backprop is only used in training, and operates within a machine
learning framework (see Chapter 9). To begin, a dataset containing a
large number of example inputs and associated outputs is assembled.
This dataset is split into three. A large training set is used to determine
the best parameter values. A smaller validation set is put aside to assess
performance and guide training. A test set is dedicated to measuring the
accuracy of the network after training is finished.
Training begins with random network parameters. It proceeds by
feeding an input from the training set into the network. The network
processes this input (forward-propagation) using the current parame-
ter vales to produce an output. This network output is compared with
the desired output for that particular input. The error between the
actual and desired output is measured as the average difference between
the actual and desired outputs squared.
Imagine that the network has two classification outputs: circle and
triangle. If the input image contains a circle then the circle output
should be 1, and the triangle output 0. Early in the training process,
the network probably won’t work at all well, since the parameters are
random. So, the circle output might have a value of 23 and the triangle 13 .
The error is equal to the average of (1 − 23 ) squared and (0 − 13 ) squared,
that is 29 .
The parameters in the network are then updated based on the error.
The procedure commences with the first weight in the first neuron of
the output layer. The mathematical relationship between this partic-
ular weight and the error is determined, and this relationship is used
to calculate how much the weight should change to reduce the error
to zero. This result is reduced by a constant value, called the learning
rate, and subtracted from the current weight. The weight adjustment
has the effect of reducing the error if the network is presented with the
same input again. Multiplication by the learning rate ensures that the
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

How to Train a Brain 191

adjustment is gradual and averaged over a large number of examples.

These steps are repeated for all of the parameters in the network,
moving backwards through the layers.
Error calculation, backprop, and parameter update are performed
for every input–output pair in the training dataset. Over many iter-
ations, the error is gradually reduced. In effect, the network learns
the relationships between the input examples and the desired output
classes. Training ends when no further reduction in error is observed (a
summary of the algorithm is provided in the Appendix).
An ANN’s great strength lies in its ability to learn and generalize.
That is, the network determines the correct class for inputs that
it has never seen before, but which are similar to those that it was
trained on. In other words, a network trained with many drawings
of circles, will correctly classify a sketch of a circle that it has never
seen before. The neural network doesn’t just memorize the training
data—it learns the general relationship between the inputs and the
output classes.
Backprop enabled researchers to efficiently train multilayer networks
for the first time. As a result, networks became more accurate and
capable of more complex classification tasks. By the end of the decade,
it was proven, at least in theory, that a sufficiently large multilayer
network could learn any input-output mapping. Minsky and Papert’s
objections had been overruled.
Be that as it may, the benefits of ANNs were not obvious to most ob-
servers. Even with backprop, networks were still small. The computers
of the day just weren’t up to performing the vast number of calculations
needed to train a large network. ANNs were a mere sideshow for the
next twenty years (1986–2006). Sure, they might be a useful aid to
cognitive scientists attempting to comprehend the workings of the
brain, but they weren’t something for serious computer scientists and
electronic engineers. Proper classification algorithms relied on rigorous
mathematics and statistics, not hocus-pocus.
Amidst widespread scepticism, a few success stories hinted at the
potential of ANNs. One of the highlights of the period was a piece of
work involving Yann LeCun (Figure 11.5), the same Yann LeCun who,
as a PhD student, had presented backprop at an obscure conference in
Paris in 1985.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

192 Mimicking the Brain

Figure 11.5 Artificial neural network innovator Yann LeCun, 2016. (Courtesy
Facebook.)

Recognizing Digits
LeCun was born in Paris in 1960. He received the Diplôme d’Ingénieur
degree from the École Supérieure d’Ingénieurs en Electrotechnique et
Electronique (ESIEE) in 1983. In his second year, he happened upon a
philosophy book discussing the nature versus nurture debate in child-
hood language development. Seymour Papert was one of the contrib-
utors. From there, he found out about Perceptrons. He started reading
everything he could find on the topic. Pretty soon, LeCun was hooked.
He specialized in neural networks for his PhD at Université Pierre et
Marie Curie (1987). After graduation, LeCun worked for a year as a
post-doctoral researcher in Geoffrey Hinton’s lab at the University of
Toronto in Canada. By then, Hinton was an established figure in the
neural network community, having been a co-author on the backprop
letter published in Nature. A year later, LeCun moved to AT&T Bell
Laboratories in New Jersey to work on neural networks for image
processing.
LeCun joined a team that had been working on building a neural
network to recognize hand written digits. Conventional algorithms
weren’t at all good at this—there was too much variability in the
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Recognizing Digits 193

writing styles. Casually written 7s are easily confused with 1s, and
vice versa. Incomplete 0s can be interpreted as 6s, and 2s with long
tails are often mixed up with truncated 3s. Rule-based algorithms just
couldn’t cope.
The team acquired a large dataset of digital images by scanning
the zip codes from the addresses on envelopes passing through the
Buffalo, New York, post office. Every letter yielded five digits. In the end,
the dataset incorporated 9,298 images. These examples were manually
sorted into ten classes corresponding to the ten decimal digits (0 to 9).
The team developed an ANN to perform the recognition task but
had little success with it. The complex mapping necessitated a large net-
work. Even with backprop, training the network had proven difficult.
To solve the problem LeCun suggested an idea that he had tinkered with
in Hinton’s lab.
The ANN took a 16x16 pixel greyscale image of a single digit as input.
The network output consisted of ten connections—one for each digit
class between 0 and 9. The output with the strongest signal indicated
the digit recognized.
LeCun’s idea was to simplify the network by breaking it up into
lots of small networks with shared parameters. His approach was to
create a unit containing just twenty-five neurons and a small number
of layers. The input to the unit is a small portion of the image—a
square of 5x5 pixels (Figure 11.6). The unit is replicated sixty-four times
to create a group. The units in the group are tiled across the image, so

Figure 11.6 Grayscale image of the digit 7 (16×16 pixels). The pixel inputs to
the unit at the top left are highlighted. Sixty-four copies of this unit are spread
over the image.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

194 Mimicking the Brain

that it is entirely covered in units. Every unit’s input overlaps with its
neighbour’s by three pixels.
The overall network contains twelve groups. Since the units in a
group share the same parameters, all perform the same function but
are applied to a different part of the image. Each group is trained to
detect a different feature. One group might detect horizontal lines in
the picture, another vertical, still another diagonal. The outputs from
each group are fed into a fully connected three-layer network. These
final layers fuse the information coming from the groups and allow
recognition of the digit in its entirety.
The network is hierarchical in structure. A single unit detects a 5×5
motif in the image. A group spots a single motif anywhere in the image.
The twelve groups detect twelve different motifs across the image. The
final, fully connected layers detect the spatial relationships between the
twelve motifs. This hierarchical organization draws inspiration from the
human visual cortex, wherein units are replicated and successive layers
process larger portions of the image.
The beauty of LeCun’s scheme is that all of the units in a single group
share the same weights. As a consequence, training is greatly simplified.
Training the first layers in the network only involves updating twelve
units, each containing just twenty-five neurons.
The mathematical process of replicating and shifting a single unit
of computation across an image is called convolution. Hence, this type of
network became known as a convolutional neural network.
The Bell Labs convolutional neural network proved to be extremely
effective, achieving a breathtaking accuracy of ninety-five per cent.
The network was close to human accuracy. The team’s findings were
published in 1989 and the system commercialized by AT&T Bells Labs.
It was estimated that, in the late 1990s, ten to twenty per cent of bank
cheques signed in the US were automatically read by a convolutional
neural network.
In 2003, Yann LeCun left Bell Labs to be appointed Professor of Com-
puter Science at New York University. Meanwhile, Geoffrey Hinton, his
old mentor at the University of Toronto, was putting together a posse.

Deep Learning
Hinton (Figure 11.7) was born in post-war Wimbledon, England (1947).
Hinton reckons that he wasn’t particularly good at maths in school.
Nevertheless, he gained entry to the University of Cambridge, enrolling
in Physics and Physiology. Dissatisfied, he transferred to Philosophy.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Deep Learning 195

Figure 11.7 Deep neural network pioneer Geoffrey Hinton, 2011. (Courtesy
Geoffery Hinton.)

Finally, he settled on Psychology. Looking back, Hinton says that he

wanted to understand how the human mind works. He concluded that
the philosophers and psychologists didn’t have the answers. He turned
to computer science.
On graduation, the young man worked as a carpenter for a year
before embarking on a PhD at Edinburgh University. With the grudging
acquiescence of his supervisor, Hinton persisted in pursing research on
ANNs. On completion of his doctorate, Hinton wended the itinerant
path of the fledgling academic. He worked at the University of Sussex,
the University of California-San Diego, Carnegie Mellon University,
and University College London (UCL) before joining the University of
Toronto as a Professor.
In 2004, Hinton submitted a proposal requesting funding for a re-
search project focused on neural computation to the Canadian Insti-
tute For Advanced Research (CIFAR). CIFAR was known for funding
basic research, but it was still a long shot. Yoshua Bengio, Professor at
the Université de Montréal, later commented: 223
It was the worst possible time. Everyone else was doing something differ-
ent. Somehow, Geoff convinced them.
The modest grant paid for a series of invite-only meetups involving
some of the top ANN researchers in the world. Bengio again: 223
We were outcast a little bit in the broader machine learning community:
we couldn’t get our papers published. This gave us a place where we could
exchange ideas.
The grant turned out to be the beginnings of a tectonic shift.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

196 Mimicking the Brain

In 2006, Hinton, Simon Osindero, and Yee-Whye Teh (University of

Toronto, National University of Singapore) published a transformative
paper. The document marked the beginnings of what is now known
as deep learning. The paper described a network made up of three fully
connected hidden layers. The network had so many parameters that
training by means of backprop was prohibitively slow. To solve the
problem, Hinton and team devised a novel procedure to accelerate
training.
Normally, backprop starts with random parameter values. Instead,
the team inserted a pretraining stage before backprop. The purpose of
this new stage was to quickly find a good set of parameters from which
backprop could start.
Backprop is an example of supervised training. This means that the
network is provided with examples of matching inputs and outputs.
In this new preliminary stage, Hinton and his co-authors proposed
employing unsupervised training. Unsupervised training only uses input
examples.
During unsupervised pre-training, example inputs are fed into the
network. The network parameters are adjusted by an algorithm so that
the ANN learns to detect significant patterns in the input. The network
isn’t told what classes these patterns are associated with—it just learns
to distinguish the patterns. For handwriting recognition, these patterns
might be the length and direction of lines or the position and length of
curves. To achieve this, the training algorithm updates the parameters
for just one layer at a time, starting with the input layer. In other words,
the algorithm grows the network parameters from the input forward.
This approach has significantly lower computational complexity than
backprop.
Once pretraining is complete, the network is able to distinguish the
most prominent patterns in the input dataset. After that, supervised
training is applied as normal, commencing with the pretrained param-
eters. Since backprop has a good starting point, it requires far fewer
iterations to complete training.
Following in the footsteps of Bell Labs, Hinton’s team elected to
tackle hand written digit recognition. This time, a much larger dataset
was available. The project used the MNIST dataset amassed by LeCun,
Corinna Cortes (Google Labs), and Christopher Burges (Microsoft Re-
search). MNIST contains 70,000 hand written digits culled from US
census returns and high school exam scripts.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Tsunami 197

The resulting ANN achieved an accuracy of 89.75 per cent, which was
not as good as LeCun’s convolutional neural network. However, that
wasn’t the point. They had proven that, by means of pretraining, a deep,
fully connected network could be trained. The road to deeper and more
effective networks was open.
Over the course of the next decade, deep learning gained mo-
mentum. The confluence of three advances enabled researchers
to build larger and deeper networks. Smarter algorithms reduced
computational complexity, faster computers reduced run-times, and
larger datasets allowed more parameters to be tuned.
In 2010, a team of researchers in Switzerland conducted an exper-
iment to see if increasing the depth of a neural network really did
translate into improved accuracy. Led by long-time neural network
guru Jürgen Schmidhuber, the group trained a six-layer neural network
to recognize digits. Their network contained a whopping 5,710 neurons.
They, like Hinton’s group, used the MNIST dataset of hand written
digits. However, even MNIST wasn’t big enough for Schmidhuber’s
team’s purposes. They artificially generated additional digit images by
distorting the MNIST photographs.
The resulting ANN achieved an accuracy of 99.65 per cent. This
wasn’t just a world record, this was human-level performance.
Suddenly, it dawned on everyone that ANNs had been too small to
be of any practical use. Deep networks were the way to go. A revolution
in artificial intelligence was at hand.

The Tsunami
The deep learning tsunami hit in three waves: first, speech recognition,
then image recognition, next natural language processing. Half a
century of pattern recognition research was swept away in just
three years.
For sixty years, the tech community had struggled to accurately
convert spoken words to text. The best algorithms relied on the
Fourier transform (see Chapter 2) to extract the amplitude of
the harmonics. Hidden Markov Models (HMMs) were then used
to determine the phonemes uttered based on the observed har-
monic content and the known probability of sound sequences in
real speech.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

198 Mimicking the Brain

With the help of Navdeep Jaitly, an intern from Hinton’s Lab, Google
ripped out half of their production speech recognition system and
replaced it with a deep neural network. They resulting hybrid ANN–
HMM speech recognition system contained a four-layer ANN. The team
trained the ANN with 5,870 hours of recorded speech sourced from
Google Voice Search, augmented with 1,400 hours of dialogue from
YouTube. The new ANN–HMM hybrid outperformed Google’s old
HMM-based speech recognition system by 4.7 per cent. In the context
of automatic speech recognition, this was a colossal advance. With
his mission at Google accomplished, Jaitly—intern extraordinaire—
returned to Toronto to finish his PhD.
Over the course of the next five years, Google progressively extended
and improved their ANN-based speech recognition system. By 2017,
Google’s speech recognition system had attained ninety-five per cent
accuracy—a previously unheard-of level accuracy.
In 2012, Hinton’s group reported on a deep neural network designed
to recognize real-world objects in still images. The objects were every-
day items such as cats, dogs, people, faces, cars, and plants. The problem
was a far cry from merely recognizing digits. Digits are made up of lines,
but object identification requires analysis of shape, colour, texture, and
edges. On top of that, the number of object classes to be recognized
greatly exceeded the paltry ten Hindu–Arabic digits.
The network—dubbed AlexNet after lead designer Alex Krizhevsky—
contained 650,000 neurons and sixty million parameters. It incorporated
five convolutional layers followed by three fully connected layers.
In addition, the work introduced a simple, yet surprisingly effective,
technique. During training, a handful of neurons are selected at
random and silenced. In other words, they are prevented from firing.
Drop-out, as the technique was named, forces the network to spread the
decision-making load over more neurons. This has the effect of making
the network more robust to variations in the input.
The team entered the network into the ImageNet Large Scale Vi-
sual Recognition Challenge in 2012. The dataset for the competition
consisted of approximately 1.2 million training images and 1,000 object
classes. Krizhevsky, Ilya Sutskever, and Hinton’s deep convolutional
network swept the boards. AlexNet achieved a top five accuracy of 84.7
per cent. That is to say, the true object class was among the ANN’s top
five picks more than 84 per cent of the time. The network’s error rate
was almost half that of the second placed system.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Tsunami 199

Meanwhile, just 500 km east along the St. Lawrence River from
Toronto, a team at the Université de Montréal was investigating how
deep neural networks could be applied to the processing of text. That
team was led by Yoshua Bengio (Figure 11.8).
Hailing from Paris, France (born 1964), Bengio was one of the lead-
ing lights of the neural network renaissance. He studied Electronic
Engineering and Computer Science at McGill University in Montreal,
obtaining BEng, MSc, and PhD degrees. A science fiction fan as an
adolescent, Bengio became passionate about neural network research
as a graduate student. He devoured all of the early papers on the topic.
A self-professed nerd, he set out to build his own ANN. After working
as a post-doctoral researcher at AT&T Bell Labs and MIT, Bengio joined
the Université de Montréal as a faculty member in 1993. Bengio’s team
trained ANNs to predict the probability of word sequences in text.
In 2014, Google picked up on Bengio’s work and adapted it to the
problem of translating documents from one language to another. By
then, the Google Translate web service had been in operation for eight
years. The system relied on conventional approaches to segment sen-
tences and map phrases from one language to another. On the whole,

Figure 11.8 Neural network researcher Yoshua Bengio, 2017. (© École polytech-
nique - J. Barande.)
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

200 Mimicking the Brain

the system’s translations weren’t particularly good. The translated sen-

tences were mostly readable but hardly fluent.
Google took the unusual step of connecting two neural networks
back to back. In the scheme, the output of the first network—the
encoder—feeds into the output of the second—the decoder. The idea was
that the encoder would convert text written in English to an abstract
vector of numbers. The decoder would then reverse the process, con-
verting the abstract vector of numbers into French. The researchers
didn’t specify the intermediate number vector. They simply relied on
the training procedure to find a suitable representation.
After two years of effort, Google completed development of an eight-
layer encoder and a matching eight-layer decoder. The network was
trained on a corpus of thirty-six million manually translated sentence
pairs. The new system outperformed the previous Google Translate
production system, reducing the number of translation errors by an
impressive sixty per sent. When the system went live on the Google web
site, bilingual users reported a sudden and dramatic improvement in
translation quality.
Success after success bred a deep learning stampede. Companies
foresaw a plethora of new applications powered by deep learning—
self-driving cars, smart cameras, next-generation recommenders, en-
hanced web search, accurate protein structure prediction, expedited
drug design, and many more. Google, Facebook, IBM, Apple, Amazon,
Yahoo!, Twitter, Adobe, and Baidu snapped up deep learning talent.
Rumours abounded of seven-figure starting salaries for the neural
network rock stars. LeCun was appointed Director of AI Research at
Facebook. Andrew Ng joined Baidu as Chief Scientist. At the age of 65,
Geoffrey Hinton became a Google summer intern!
In 2015, amidst the gold rush, LeCun, Hinton, and Bengio published
a paper in Nature surveying developments. Deep neural networks had
swept the entire field of artificial intelligence before it. Everything was
changed, utterly.
LeCun, Hinton, and Bengio were the recipients of the 2018 ACM
Turing Award, sharing the Google sponsored $1 million prize.
With the runaway success of deep learning, some have speculated
that human-level artificial general intelligence (see Chapter 5) is just
around the corner. LeCun demures: 236
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Tsunami 201

Whether we’ll be able to use new methods to create human-level intelli-

gence, well, there’s probably another fifty mountains to climb, including
ones we can’t even see yet. We’ve only climbed the first mountain. Maybe
the second.
All we have to date are sophisticated pattern recognition engines. Yet,
we may speculate on the path that might take us across these moun-
tains. Presently, the best guess is that a network of ANNs will be re-
quired. Significant improvements may also require a fundamental re-
working of the ANN. Today’s ANNs are only a rough approximation
of what goes on in a biological neural network. It may be that a more
realistic model is needed. The devil may well be in the details.
For those outside the computer science community, the first hint
of the power of deep neural networks came in 2016. In that year, an
artificial intelligence hit the world’s news media headlines. Albeit in a
narrow field of endeavour, an artificial intelligence had, perhaps for the
first time, attained superhuman ability.
OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

12
Superhuman Intelligence

Because previous moves cannot be changed, subsequent regrets

are truly difficult to bear.
Unknown author, translated by Robert W. Foster
The Classic of Go, 6th century 237

19 March 2016. A young man walks purposefully down the corridors of

Seoul’s Four Seasons Hotel. Along the way, he passes ranks of journalists
and photographers clamouring for his attention. Dressed in a sharp
navy suit and open-necked shirt, he looks much younger than his
thirty-three years. The man is thin and Asian in appearance. His hair
is combed from the crown of his head towards an unwavering fringe.
His upper lip bears the ghost of a moustache. Despite all the attention,
he seems relaxed and confident.
The man leaves the cacophony of the corridors behind and enters a
hushed conference room. A small audience and a handful of television
cameras face a low, neon blue stage. The man settles into a black leather
chair set on the right of a low podium. The lettering on the podium
indicates that this is ‘Lee Sedol’. A South Korean flag confirms his
nationality.
Opposite Lee is Aja Huang. The lettering on Huang’s side of the
podium spells out one word: ‘AlphaGo’. Although Huang is Taiwanese,
the flag below is British. A computer monitor, keyboard, and mouse
rest alongside Huang. A panel of judges sits behind the two men,
overlooking them. A table separates Lee and Huang. The table bears a Go
board, a timer, and four bowls. Two of the bowls are empty. One holds
a collection of identical white stones, the other similar black pieces.
Lee is widely acknowledged to be one of the top five Go players in
the world. He has won eighteen international titles. A child prodigy,
he studied at the famous Korean Baduk (Go) Association. Lee turned
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

204 Superhuman Intelligence

professional at just twelve years of age. Among Go aficionados, he has a

reputation for aggressive and imaginative play. Lee Sedol—the ‘Strong
Stone’—is a national figure in South Korea.
AlphaGo is a computer program. The codification of a complex
game-playing algorithm. Huang’s job is to relay Lee’s moves to AlphaGo
and to place stones on the board for the computer. AlphaGo is the
creation of a small London-based outfit by the name of DeepMind
Technologies. Huang is lead programmer on the AlphaGo project. Two
years previously, DeepMind was acquired by Google for an estimated
$500–625 million.
Google is the sponsor of the AlphaGo–Lee match. The prize money
is set at $1 million. If the computer wins, the money goes to charity.
In the days leading up to the match, Lee was confident. At one press
conference, he claimed that the question wasn’t whether he would
win the match. Rather, it was whether he would lose a single game.
Lee seemed justified in his confidence. A computer had never beaten
a top 300 professional Go player in a competitive match. Prior to the
showdown in Seoul, grandmasters predicted that Lee would pocket an
easy million bucks.
Like Chess, Go is an abstract war simulation. The game originated in
China around three thousand years ago and spread to Korea and Japan
during the fifth to seventh centuries CE. Go remains hugely popular in
east Asia.
The Championship version of the game is played on a 19 × 19 grid.
At the start, the grid is empty. Players take turns to place one stone
on a grid intersection. An intersection is called a territory. A player can
pass their turn, if they so wish. One player places black stones, the
other white. The aims of the game are to enclose territory and capture
enemy stones. A player’s stones are removed from the board when they
are encircled by their opponent’s pebbles. The game ends when both
players pass consecutively rather than play moves. A player can resign
by placing a stone off-grid. The player with the greatest number of
territories plus captives wins. The player that is second to start is given
a small points bonus in compensation. In competitive matches, moves
are made on a timer.
Watching a video of Go on fast-forward video is hypnotic. Complex
patterns of black and white stones evolve, coalesce, and colonize the
board in a slow dance. Sudden bursts of activity transform the board as
stones are encircled and vanish. Go devotees see an underlying beauty
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Match 205

in the game. To them, a match is a reflection of a player’s imagina-

tion, courage, and fortitude. The values of the game—elegance and
humility—are instilled in players from an early age.
Although the rules are simple to learn, Go is highly complex to play.
A Go board is more than five times the size of a Chess board (8x8).
Go games last, on average, 150 moves. In making a move, a Go player
must consider around 250 possibilities. The theoretical Go game tree
(see Chapter 5) contains an astronomical 250150 , or 10359 , nodes. By this
estimate, Go is 10226 times (a one with 226 zeros after it) more complex
than Chess.

The Match
The AlphaGo-Lee match is a best-of-five games contest. An estimated
sixty million television viewers are watching the game on television
in China alone. A hundred thousand enthusiasts are glued to the live
English language coverage on YouTube.
The DeepMind team spectates from a war room in the bowels of the
hotel. The room is kitted out with a wall of monitors. Some screens dis-
play camera feeds from the match room. Others show lists of numbers
and graphs summarizing AlphaGo’s analysis of the game. DeepMind
CEO Demis Hassabis and lead project researcher Don Silver watch the
match unfold from this vantage point. Like the rest of their team,
Hassabis and Silver are anxious, but powerless.
Day one, game one. Lee places the first stone. Bizarrely, it takes
AlphaGo half a minute to respond. The AlphaGo team holds its breath.
Is the machine working at all? Finally, it makes its decision and Huang
places AlphaGo’s first stone.
AlphaGo attacks from the outset. Lee seems mildly surprised. Al-
phaGo isn’t playing like a computer at all. Then comes AlphaGo’s move
102. It is aggressive—a gateway to complicated skirmishes. Lee recoils,
rubbing the back of his neck. He looks worried. He strengthens his
resolve and rejoins battle. Eighty-four moves later, Lee resigns. The
reaction in the DeepMind team room is euphoric.
Afterwards, Lee and a composed Hassabis face the assembled media
at the post-game press conference. The two sit apart on stools on a bare
stage. Lee looks isolated, lost, abandoned. He is deeply disappointed but
accepts his loss with grace. The following morning, AlphaGo’s victory
is front page news.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

206 Superhuman Intelligence

Day two, game two. This time, Lee knows what to expect. He plays
more cautiously. On move 37, AlphaGo makes an unexpected play—
a move that humans seldom play. In shock, Lee walks out of the
conference room. Huang and the match judges stay put, bewildered.
Minutes later, having collected his thoughts, Lee returns to the fray.
After 211 moves, Lee again resigns.
AlphaGo’s move 37 was decisive. The computer estimated that the
chances of a human playing the move was one in ten thousand. The
European Go champion Fan Hui was awestruck. For him, move 37
was, ‘So beautiful. So beautiful.’ AlphaGo had displayed insight beyond
human expertise. The machine was creative.
At the press conference, Lee reflected on the game: 242
Yesterday, I was surprised. But today I am speechless. If you look at the
way the game was played, I admit, it was a very clear loss on my part.
From the very beginning of the game, there was not a moment in time
when I felt that I was leading.

Day three, game three. Lee’s facial expressions say it all—initial calm,
turning to concern, followed by agony, and finally dismay. He resigns
after four hours of play. Against all expectations—save those of Google
and DeepMind—AlphaGo wins the match.
Lee looks worn out. Regardless, he is gracious in defeat: 243
I apologize for being unable to satisfy a lot of people’s expectations.
I kind of felt powerless.
A strange kind of melancholy descends on proceedings. Everyone is
affected, even the DeepMind team. Those present are witnessing the
suffering of a great man. One of Lee’s rivals remarks that Lee had
fought: 238
A very lonely battle against an invisible opponent.

Even though the series is decided, Lee and AlphaGo press on, playing
games four and five. In game four, Lee is more himself. The Strong Stone
takes a high-risk strategy. His move 78—a so-called ‘wedge’ play—is
later referred to by commentators as ‘God’s move’. AlphaGo’s response
is disastrous for the machine. Soon its play becomes rudderless. Finding
no way out, the computer begins to make nonsense moves. Eventually,
AlphaGo resigns.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Winning Move 207

Lee tries the same high-risk approach in game five. This time, there is
no miracle play. Lee is forced to resign.
AlphaGo wins the match by 4 games to 1.

The Winning Move

AlphaGo’s victory sent shockwaves through both the Go and the com-
puter science communities. Based on projections of computer perfor-
mance, this wasn’t supposed to happen for at least another fifteen years.
The theory was that tackling Go needed far faster hardware than was
available in 2016. In reality, the secret to AlphaGo’s success lay in its
algorithms, not in its hardware.
The AlphaGo hardware was mundane by the standards of 2016.
During development, the DeepMind team used only forty-eight central
and eight graphics processing units—something that a hobbyist could
easily rig together in their garage. In competition, AlphaGo ran on
computers in one of Google’s Internet-connected data centres. The
program occupied 1,920 central and 280 graphics processing units. In
contrast, the most powerful supercomputer at the time—the Chinese
Tiamhe-2—had 3.1 million central processing units. AlphaGo was a
slouch by comparison.
Like Arthur Samuel’s Checkers playing program, the AlphaGo algo-
rithm utilizes Monte Carlo tree search (see Chapter 5). On the com-
puter’s turn, it hunts for the most promising next play. For each of
these, it examines the mostly likely responses by its opponent. Follow-
ing that, it evaluates its own possible replies. In this way, the computer
produces a tree of possible future plays with the current board position
as its root.
Once the tree has been built, the computer uses a minimax procedure
to select the best move (see Chapter 5). The computer starts with the
furthest look-ahead board positions—the leaves of the tree. It then
moves backwards through the tree towards the root. At every branch-
ing point, it propagates the best move backwards in the tree. On its own
plays, the best choice is the one that maximizes the computer’s chances
of a win. On its opponent’s plays, the computer selects the play that
minimizes its own probability of winning. When the procedure reaches
the root of the tree, the computer selects the play which, it considers,
will give it the best chance of winning the game in the long term.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

208 Superhuman Intelligence

AlphaGo uses ANNs to evaluate board positions. A board position

is represented by a table of numbers. Each number indicates whether
there is a black stone, a white stone, or no stone at a grid intersect. To
evaluate a position, the table of numbers is input to an ANN. The neural
network outputs a score indicating the strength of the position.
AlphaGo’s neural networks are larger versions of the convolutional
neural networks introduced by Yann LeCun to recognize digits (see
Chapter 11). In effect, the table of digits is treated much like an image.
AlphaGo’s neural networks recognize the patterns on the board, much
as LeCun’s network recognized the lines and curves in numerals.
AlphaGo uses three neural networks.
The first is a value network. The value network estimates the probability
of winning from a given position. The value network scores the posi-
tions at the end of the tree search.
The second ANN is a policy network. The policy network serves to guide
the tree search. The policy network scores a position based on how
promising it is. If a position looks like it might lead to a win in the future,
it is given a high policy score. Only positions with high policy scores are
investigated in greater depth. In this way, the policy network controls
the breadth of the search.
If the value network was completely accurate then tree search would
not be necessary. The computer could simply evaluate all of the next
positions and choose the best. The lookahead improves accuracy by
rolling the positions forward. As the game gets closer to the end,
it becomes easier to predict the outcome and so the value network
becomes more accurate.
Ideally, the value network would also be used for policy decisions.
Again, the value network isn’t sufficiently accurate. A separate policy
network offers greater accuracy in evaluating early game positions. The
value network is trained for precision, whereas the policy network is
trained not to miss promising paths in the tree search.
The third network is a SL-value (Supervised Learning) network. This
network is trained to score positions in the same way as humans. The
other networks seek to determine the true chances of winning. The
SL-value network allows the computer to predict the most likely move
that a human player will make.
AlphaGo’s neural networks were trained in three stages.
In stage one, the SL-value network was trained using supervised
learning (see Chapter 11). The network contained thirteen layers of
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

The Winning Move 209

neurons. Training was performed using positions and moves obtained

from the KGS Go database. KGS allows players from all over the world
to play Go online for free. Games are recorded and are available on the
KGS web site. AlphaGo used this database to supply examples of board
positions and the moves played by humans. Thirty million positions
from 160,000 games were used to train the network.
In stage two, the SL-value network was refined to create a policy net-
work. This time, reinforcement learning was used. The network played Go
against itself. The reinforcement algorithm used the outcome of every
game (win or lose) as a reference with which to update the network
parameters. AlphaGo played 1.2 million games against a pool of older
versions of itself. As it played more games, AlphaGo’s performance
gradually edged upwards. In trials, the resulting policy network beat
the original SL-value network in eighty per cent of games.
In stage three, the team used the policy network to seed a value
network. Again, reinforcement learning was used. Rather than play-
ing entire games from scratch, intermediate positions from the KGS
database were used as starting points. A further thirty million games
were played to complete training of the value network.
These three neural networks were the chief difference between Al-
phaGo and prior Go playing computers. Previously, positions were
evaluated using hand-crafted rules and scoring methods (i.e. expert
systems and case-based reasoning). AlphaGo’s ANNs offered far more
accurate position evaluation.
The accuracy of AlphaGo’s ANNs comes from the confluence of
three factors. First, deep ANN are extremely good at learning complex
relationships between inputs and outputs. Second, during training,
the networks were exposed to huge volumes of data. In preparation
for the match, AlphaGo inspected more Go moves than any human,
ever. Third, advances in algorithms and hardware allowed these large
networks to be trained in a reasonable time frame.
These factors explain why AlphaGo outperformed previous pro-
grams, but they do not explain why AlphaGo beat Lee Sedol. Analysis
of Go and Chess grandmasters thought processes suggest that they
evaluate far fewer positions than AlphaGo. The human tree search is
much narrower and shallower than the computer’s. Human pattern
recognition must, therefore, be much more effective than AlphaGo’s.
AlphaGo makes up for this deficiency by means of faster processing.
In competition, the machine evaluates many more board positions than
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

210 Superhuman Intelligence

humans. The high speed of AlphaGo’s electronic simulations of neuron

behaviour allowed it examine more positions in the tree search. That is
how AlphaGo beat Lee Sedol.
The implication was that there remained plenty of scope to improve
the pattern recognition capabilities of ANNs.

DeepMind
To outsiders, it must have seemed that DeepMind was an overnight
success but, of course, it wasn’t. Demis Hassabis, the company’s co-
founder and CEO, had been thinking about board games and computers
since he was a kid.
Hassabis (Figure 12.1) was born in London, England, in 1976. He is
proudly ‘North London born and bred’. 245 Hassabis reached master
level in Chess aged 13. He spent his winnings on his first computer—
a Sinclair Spectrum 48K—and taught himself to program. Before long,
he had completed his first Chess-playing program.
Hassabis finished high school at sixteen and joined a video game
development company (Lionhead Studios). A year later, he was co-
designer and lead-programmer on the popular management simula-
tion game Theme Park. Hassabis left the company to enrol in a Computer

Figure 12.1 DeepMind co-founder and CEO Demis Hassabis, 2018. (Courtesy
DeepMind.)
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

DeepMind 211

Science degree programme at Cambridge University. In his spare time,

he entered the Pentamind competition at the annual Mind Sports
Olympiad. Pentamind pits elite players against each other across five
board games: Backgammon, Chess, Scrabble, Go, and Poker. Hassabis
went on to win the competition a record five times.
Reflecting on the range of his accomplishments, Hassabis says: 246
I get bored quite easily, and the world is so interesting, there are so many
cool things to do. If I was a physical sportsman, I’d have wanted to be a
decathlete.
After graduation from Cambridge, Hassabis founded his own inde-
pendent video games development company. Elixir Studios released
two games but then hit problems. The firm was wound up in 2005.
Hassabis’s formal statement on the closure belied his disappointment
at the turn of events: 247
It seems that today’s games industry no longer has room for small inde-
pendent developers wanting to work on innovative and original ideas.
Hassabis resolved that his career should take a new direction. He set
out on a mission to build artificial intelligence. Believing that the best
first step lay in understanding how biological intelligence works, he
embarked on a PhD in Cognitive Neuroscience at UCL. The subject ex-
plores how the human brain works, often employing computer models
to better comprehend brain function. Hassabis published a string of
important research papers on the topic before graduating.
Armed with these fresh insights, Hassabis co-founded DeepMind
Technologies with Shane Legg and Mustafa Suleyman in 2010. Hassabis
and Suleyman had known each another since childhood. Hassabis and
Legg met while studying for their PhDs at UCL .
DeepMind first came to the attention of the wider scientific com-
munity thanks to a letter published in Nature magazine. The letter
described an artificial neural network that DeepMind trained to play
Atari video games. Atari video games are the coin-operated classics from
the arcades of the 1980s, including Space Invaders, Breakout, and River
Raid. DeepMind’s neural network took the screen image as input. Its
output was the control signals for the game—the joystick twitches and
the button presses. In effect, the ANN replaced the human player.
DeepMind’s ANN taught itself how to play Space Invaders from
scratch. Its sole preprogrammed aim was to score as many points as
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

212 Superhuman Intelligence

possible. At first, the network played randomly. Through trial and error
and a learning algorithm, it gradually accumulated a suite of point-
scoring tactics. By the end of training, DeepMind’s neural network was
better at Space Invaders than any previous algorithm. This, in itself, was
an achievement. What was remarkable was that the network went on
to learn how to play forty-nine different Atari video games. The games
were varied, requiring different skills. Not only could the network play
the games, it could play them just as well as a professional human games
tester. This was new. DeepMind’s ANN had excelled across a range
of tasks. For the first time, an ANN was showing a general-purpose
learning capability.
A year later—and just two months before the Lee match—
DeepMind published another paper in Nature. In it, they described
AlphaGo and casually mentioned that the program had beaten the
European Go champion Fan Hui. The paper should have been a
warning to Lee Sedol and others. However, Europe was regarded as
a Go backwater. It was presumed that Fan Hui had erred. Fan Hui, for
his part, was so impressed with AlphaGo that he accepted an offer to
act as a consultant to the DeepMind team as they prepared for the Lee
Sedol match.
AlphaGo’s resounding victory over Lee garnered headlines world-
wide. In contrast, AlphaGo’s later defeat of the world number one was
an anticlimax. AlphaGo beat nineteen-year-old Ke Jie 3–0 in May 2017.
This time, the match received little media coverage. The world seemed
to have accepted humankind’s defeat and moved on. After the win,
Hassabis said that, based on AlphaGo’s analysis, Jie had played almost
perfectly. Almost perfect was no longer good enough. After the match,
DeepMind retired AlphaGo from competitive play.
Yet the company didn’t stop working on computers that played Go. It
published another paper in Nature, describing a new neural network pro-
gram dubbed AlphaGo Zero. AlphaGo Zero employed a reduced tree
search and just one neural network. This single two-headed network
replaced the policy and value networks of its predecessor. AlphaGo
Zero used a new, more efficient training procedure based exclusively on
reinforcement learning. Gone was the need for a database of human
moves. AlphaGo Zero taught itself how to play Go from scratch in a
meagre forty days. In that time, it played twenty-nine million games.
The machine was allowed just five seconds of processing time between
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

DeepMind 213

moves. AlphaGo Zero was tested against the version of AlphaGo that
beat Ke Jie. AlphaGo Zero won 100 games to nil.
In just forty days the computer had taught itself to play Go better
than any human, ever. AlphaGo Zero was emphatically superhuman.
Human Go grandmasters pored over AlphaGo Zero’s moves. They
discovered that AlphaGo Zero employed previously unknown game-
winning strategies. Ke Jie began to include the new tactics in his own
repertoire. A new age in the history of Go was dawning. Human grand-
masters were now apprentices to the machine. Biological neural net-
works were learning from their artificial creations.
The true significance of AlphaGo Zero doesn’t lie on a Go board, how-
ever. Its real importance lies in the fact that AlphaGo Zero is a prototype
for a general-purpose problem solver. The algorithms embedded in its
software can be applied to other problems. This capability will allow
ANNs to rapidly take on new tasks and solve problems that they haven’t
seen before—something that heretofore only humans and high-level
mammals have achieved.
The first signs of this general problem-solving capacity appeared in
2018 in yet another Nature paper. This time, the DeepMind team trained
an ANN named AlphaZero to play Go, Chess, and Shogi (Japanese
Chess). It wasn’t particularly surprising to read that AlphaZero learned
how to play the three games solely from self-play. Neither was it es-
pecially eyebrow raising to note that AlphaZero defeated the previ-
ous world-champion programs (Stockfish, Elmo, AlphaGo Zero) at all
three games. What was jaw dropping was that, starting from random
play, AlphaZero learned to play Chess in just nine hours, Shogi in
twelve hours, and Go in thirteen days. The human mind was beginning
to appear weak in comparison.
OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

13
Next Steps

In terms of cryptocurrencies, generally, I can say with almost

certainty that they will come to a bad ending.
Warren Buffet
On CNBC, 2018 251

Computer algorithms have fundamentally altered the way that we

live. Information technology is deeply embedded in our workplaces.
Communication is entrusted to email, social media, and messaging
apps. Our leisure time is dominated by video games, streaming music,
and online movies. Recommenders manipulate our purchasing deci-
sions. Our romantic liaisons are prompted by algorithms. Much has
changed. Yet, many more revolutionary technologies are under devel-
opment in the hip open-plan offices of the tech giants, the makeshift
workspaces of impoverished start-ups, and the tatty labs of college
professors. In this, the final chapter, we examine two new algorithms
that have the potential to change the world.

Cryptocurrency
The first of these is the algorithm that underpins cryptocurrency. Cryp-
tocurrency is a form of money that only exists as information held in
a computer network. The world’s first cryptocurrency, Bitcoin, now
has over seventeen million ‘coins’ in circulation with a total real-world
value of $200 billion (2019). Cryptocurrencies seem set to disrupt the
global financial system.
The origins of cryptocurrency lie in the Cypherpunk movement that
began in the 1990s. The Cypherpunks are a loose amalgam of skilled
cryptographers, mathematicians, programmers, and hackers who be-
lieve passionately in the need for electronic privacy. Connected by
mailing lists and online discussion groups, the Cypherpunks develop
open-source software that users can deploy for free to secure their data
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

216 Next Steps

and communications. Their ideals were set out by Eric Hughes in A

Cypherpunk’s Manifesto (1993): 253
Privacy is necessary for an open society in the electronic age.
We cannot expect governments, corporations, or other large, faceless
organizations to grant us privacy out of their beneficence.
Privacy in an open society […] requires cryptography.
Cypherpunks write code. We know that someone has to write software
to defend privacy, and since we can’t get privacy unless we all do, we’re
going to write it. We publish our code so that our fellow Cypherpunks
may practice and play with it. Our code is free for all to use, worldwide.
The Cypherpunks contributed their skills to a series of projects
to develop secure software. PGP enabled RSA encryption for email
users. Tor allowed anonymous web browsing. The group wrote
white papers on matters of encryption. They filed lawsuits against
the US government in regard to export controls on encryption
technologies. On occasion, they even urged public disobedience in
support of their aims. The Cypherpunks also promoted the concept of
cryptocurrency.
In their eyes, cryptocurrency has three key advantages over con-
ventional currency. Firstly, a cryptocurrency is not controlled by a
central authority. There is no central bank of Bitcoin. The currency is
managed by a network of computers. Anyone can join the network.
There is no application form. Volunteers simply download the cryp-
tocurrency software from the Internet and run it. No computer on
the network is more important than any other. All are peers. Sec-
ondly, users are anonymous, provided that they do not trade cryptocoins
for conventional currency. Privacy is guaranteed by means of pub-
lic key cryptography. Anyone can be a user. They simply download
an app, which submits their transactions to the network. Thirdly,
transactions carry a low fee and zero sales tax. Furthermore crypto-
coins can be sent internationally without incurring currency exchange
charges.
While the Cypherpunks were early proponents of cryptocurrency,
no one knew how to make it work for real. There didn’t seem to be any
way around the Double-Spend Problem.
Conventional online currencies rely on a central authority that ap-
proves transactions (i.e. a transfer of money between users). The cen-
tral authority maintains a ledger of all transactions. The ledger is the
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Bitcoin 217

electronic equivalent of the pen-and-paper logs maintained by banks

100 years ago. By inspection of the ledger, the central authority knows
how much money every user has in their account. When a user requests
a transaction, the central authority can easily check if they have enough
money to cover the transaction. If they do, the transaction is accepted as
valid and recorded in the ledger. If they don’t, the transaction is rejected.
The difficulty in designing a cryptocurrency is the removal of the
central authority. The ideal is that a distributed network of computers
maintains the ledger. Every computer on the network has its own
copy of the ledger. The hard part is synchronizing these copies of the
ledger (i.e. keeping them all up to date). Communication delays are
highly unpredictable on the Internet. Computers can join and leave the
network at any time. These issues lead to the Double-Spend Problem.
Imagine that Alice only has 1.5 cryptocoins in her account. She
owes money to both Bob and Charlie. In desperation, she sends two
transactions to the network. In one, she transfers 1.5 cryptocoins
to Bob. In the other, she transfers 1.5 cryptocoins to Charlie. If she
sends the two transactions to different parts of the network, there is
a chance that one computer will accept the transfer to Bob at exactly
the same time as another device accepts the transfer to Charlie. If she
is lucky, Alice will pay both parties at the same time, double-spending
her funds.

Bitcoin
Satoshi Nakamoto announced a solution to the Double-Spend Problem
on 31 October 2008 in a white paper posted to a Cypherpunk mailing
list. The paper introduced Bitcoin, the world’s first practical cryptocur-
rency. The following January, Nakamoto released the Bitcoin source
code and the original—or genesis—Bitcoin block.
Fundamentally, bitcoins are just sequences of characters (numbers
and letters) held in a computer network. A bitcoin only has value
because people believe that it has value. Users expect that they will be
able to exchange bitcoins for goods and services at a later date. In this,
Bitcoin is no different from the banknotes in your pocket. The paper
itself has little intrinsic value. Its value derives from the expectation that
you will be able to exchange it for something of worth.
Bitcoin is reasonably straightforward to use. Users buy, sell, and
exchange bitcoins via apps. Bitcoins can be used to purchase real-world
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

218 Next Steps

goods from participating retailers. Digital brokers will happily exchange

bitcoins for old-fashioned state-controlled currency. User anonymity is
protected by public key cryptography (see Chapter 7). Prior to using
Bitcoin, a user generates a public and private key pair. They keep the
private key secret. The public key functions as the user’s ID on Bitcoin.
When a user wants to send bitcoins to another user, they create
a transaction. The transaction consists of the transaction ID, the
sender’s ID, the recipient’s ID, the amount, and the input transaction’s
IDs (Figure 13.1). The inputs to the transaction are previous transactions
in which the sender received the bitcoins that they are about to spend.
The input transactions must already have been logged in the ledger
and must not have been spent previously. Say Alice wants to send B0.5
to Bob. She references two previous transactions in which she received
B0.3 from Jack and B0.2 from Jill. She does this by including the IDs of
the previous transactions in the new transaction. The amount to be
spent and the total amount referenced must match exactly. This may
mean that the sender has to send some bitcoins back to themselves as
change.
The sender’s encryption key is used to authenticate a transaction
(Figure 13.2). Authentication ensures that the sender really did want
to transfer that bitcoin amount to the receiver. It also guarantees that
the transaction isn’t from a duplicitous third party. To enable authen-
tication, the sender appends a digital signature to the transaction. A
digital signature is the equivalent of the handwritten signature on a
paper cheque.

Figure 13.1 A Bitcoin transaction must refer to previous input transactions

with the same total value.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Bitcoin 219

Figure 13.2 Creation and verification of a digital signature.

The digital signature is created by encrypting a summary of the

transaction with the sender’s private key. Normally, a public key is used
for encryption and a private key for decryption. To generate a signature,
the process is reversed. The private key encrypts and the public key
decrypts. This means that anyone can check the signature but only the
sender can create it.
When a computer on the Bitcoin network receives the transaction,
it firstly verifies that the signature is genuine (Figure 13.2). It does this
by decrypting the signature using the sender’s public key. This gives
the transaction summary. The receiver also summarises the transaction
and compares the two versions—the decrypted and the calculated. If
they match, the transaction must be authentic. Only the real sender
could have created the digital signature since only they have the private
key. If the two versions do not match, the transaction is rejected as being
invalid.
The digital signature ensures that the transaction wasn’t tampered
with en route. Any change to the message alters its summary. As a re-
sult, the calculated summary will not match the decrypted equivalent.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

220 Next Steps

The transaction summary is calculated by means of a hashing algo-

rithm. A hashing algorithm takes a large amount of text and squeezes
it down to a shorter sequence of characters. Information is lost in the
process, but the output sequence is highly dependent on the input.
In other words, a small alteration in the original text leads to a large
random change in the output. The summary is often called the hash
of the message. The hashing function used in Bitcoin is essentially an
advanced checksum algorithm (see Chapter 7).
After authentication, the new transaction is validated. The receiving
computers check that the input transactions exist in the ledger and that
the associated monies have not been spent previously.

Blockchain
Next, the transaction is confirmed and logged. The network computers
incorporate the new transaction in a larger block of unconfirmed trans-
actions. A block is simply a group of unconfirmed transactions and their
associated data. The network computers race to add their blocks to the
ledger. The winner of the race incorporates its block in the chain of
blocks. This chain of blocks—or Blockchain—is the ledger (Figure 13.3).
It links every confirmed Bitcoin block in an unbroken sequence stretch-
ing all the way back to Nakamoto’s genesis block. The links of the chain
are formed by including the ID of the previous block in the next block.
The chain rigidly defines the order in which transactions are applied
to the ledger. The transactions in a single block are considered to have
happened at the same time. Transactions in any previous block are

Figure 13.3 The Blockchain. Blocks are chained together by block ID to

enforce the order in which the transactions are applied to the ledger. Blocks
contain a unique block ID, the previous block’s ID, a group of transactions, and
a randomly generated number.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Blockchain 221

considered to have happened at an earlier time. The winner of the race

shares its block with the entire network so that all of the ledgers are
kept up to date.
The race ensures that only one computer at a time can add a block
to the chain. Winning the race depends on luck. In most cases, there
is a significant delay between the first and second placed computers.
This delay allows time for the Blockchain update from the winner to
propagate through the network. Dead heats between first and second
are unlikely, but can occur. Provision is made for this by requiring that
six blocks are added to the chain before a block is considered confirmed.
Six dead heats in succession are, for practical purposes, impossible.
To win the race, a computer has to create a valid block. This is done
by generating a random number and appending it to a candidate group
of transactions. A hashing algorithm is then applied to the block. If
the hash is less than a pre-determined threshold, the candidate block is
considered valid. If the hash is equal to, or greater than, the threshold,
the candidate block is viewed as invalid. The network computers try
different random numbers until a valid block is formed. The first com-
puter to create a valid block wins the race and shares the block with the
rest of the network. The other machines log the transactions by adding
the block to their own copy of the ledger. All of the machines then
resume their attempts to create a new valid block using transactions
which have not yet been committed to the ledger.
Creating a valid block amounts to trying lots of random numbers.
It is impossible to predict in advance which number will pass the hash-
threshold test. Trial and error is the only way to find a suitable number.
Since producing a valid block is down to luck, any computer might be
the next winner. Thus, there is no single authority for adding blocks
to the chain. The work is distributed across the entire network of peer
computers (see the Appendix for further details).
This explains how bitcoins are exchanged between users. But how are
they created in the first place?
Every time that a network computer creates a valid block, its owner
is rewarded with bitcoins. The process of validating blocks in return
for bitcoins is called mining. Validation of the genesis block released
fifty bitcoins to Nakamoto. In the same way, the owners of network
computers are rewarded for maintaining the ledger.
Bitcoin’s early adopters were buyers and sellers in the online black
marketplaces. Bitcoin’s emergence on the Darknet was fuelled by
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

222 Next Steps

illegal drug purchases. At first, the main attraction was anonymity.

In time, legitimate organizations began to accept Bitcoin payments.
Coinbase was established in 2012 as a broker for digital currencies. In
2014, Microsoft began to accept bitcoins for online purchases of Xbox
games. The value of a bitcoin in ‘hard’ currencies became a rollercoaster.
In 2011, a single bitcoin (B1) was worth 30 US cents. On 18 December
2017, the price spiked to a staggering all-time high of $19,498.63.
At first glance, mining bitcoins appears to be a way of creating money
from nothing. All that is required is to install the Bitcoin software,
download the ledger, and start mining. However, the costs of the
computer and the electricity that it consumes are real. Estimates put
the annual global revenue from Bitcoin mining at over $8 billion with
costs of more than $3 billion (2019). The rewards for mining halve
every four years. Ultimately, around twenty-one million bitcoins will
be issued.
Bitcoin’s success has led to a slew of cryptocurrencies, most notably
Ethereum (2015) and Facebook’s Libra (2019). However, companies are
also investing heavily in the Blockchain technology that underpins Bit-
coin. Blockchain offers a secure distributed ledger, independent of the
cryptocurrency aspect. A Blockchain can track and enforce sequenc-
ing on any form of transaction. Possibilities include validating legal
contracts, maintaining online identities, recording medical histories,
verifying media providence, and tracing supply chains. Blockchain will
likely reshape operational models in a wide range of industries and may,
ultimately, prove to be more useful that Bitcoin itself.

Who Is Nakamoto?
The really curious thing about Bitcoin is that no one knows who Satoshi
Nakamoto—Bitcoin’s inventor—is. The first mention of Nakamoto
was the release of the original Bitcoin white paper. Nakamoto remained
active on the Cypherpunk mailing lists for a few years, then in 2010
Nakamoto passed control of the Bitcoin source code to Gavin Andresen.
The following April, Nakamoto declared: 260
I’ve moved on to other things.
It’s in good hands with Gavin and everyone.
Save for a handful of messages—most of which are now thought to be
hoaxes—that was the last anyone heard from Nakamoto.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Quantum Computers 223

Of course, there has been much speculation regarding Nakamoto’s

identity. The clues are scant. Nakamoto is clearly a world-class cryp-
tographer (or group of cryptographers). The Bitcoin source code is
impeccable. So Nakamoto is also an expert coder. Nakamoto’s written
English is perfect. Thus it may be that Nakamoto is a native speaker.
Closer inspection of Nakamoto’s posts reveals hints of a British or Aus-
tralian accent. The genesis block includes a headline from the London
Times. Perhaps Nakamoto is from the UK. Analysis of timestamps shows
that Nakamoto mostly posted bulletin board messages between 3 pm
and 3 am Greenwich Mean Time. If this habit was due to a nocturnal
sleeping pattern, then this means that Nakamoto was likely living on
the east coast of America. There are no clues as to Nakamoto’s gender.
Lists of suspects have been bandied around online. The names of most
top-notch Cypherpunks have been suggested at one time or another.
Some individuals have even claimed to be Nakamoto. To date, however,
no one has proven that they are Nakamoto. All that has to be done
to settle the case is to decrypt a message sent using Nakamoto’s public
key. The person, or persons, that can do that must be in possession of
Nakamoto’s private encryption key.
Nakamoto’s 1.1 million bitcoins remain untouched. Presently,
1.1 million bitcoins are worth in excess of $11 billion (2019). By that
reckoning, Nakamoto is one of the richest 150 people in the world.
Why doesn’t he, she, or they step forward to claim their rightful riches?
Is their reluctance simply down to strict observation of the Cypherpunk
code of honour? Or is something more sinister going on?
Satoshi Nakamoto remains an enigma.

Quantum Computers
Bitcoin draws heavily on RSA public key cryptography to ensure user
anonymity and provide transaction authentication. In turn, the secu-
rity of the RSA algorithm hinges on the assumption that there is no fast
algorithm for prime factorization of large numbers (see Chapter 7). In
other words, there is no fast method for determining which two prime
numbers were multiplied to produce a given large number. Clearly, the
prime factors of 21 are 3 and 7, but this determination is only quick
because 21 is small. Prime factorization of a large prime can take decades
on a supercomputer.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

224 Next Steps

Bitcoin, and the whole edifice of Internet security, depends on this

single assumption of slow prime factorization. If a fast prime factor-
ization algorithm were to be invented, Bitcoin and almost every secret
message on the Internet would suddenly become vulnerable to attack.
In 1994, the spectre of just such an algorithm came into view. The only
saving grace was that the miracle algorithm required a new type of
computer. A device called a quantum computer.
In 1981, Richard Feynman delivered the keynote address at a confer-
ence at MIT. By then, Feynman was sixty-three and widely regarded
to be one of the greatest physicists of all time. During the war, he
worked on the Manhattan Project at Los Alamos. At Cornell Univer-
sity, he made huge strides in quantum electrodynamics. While at Cal-
tech, Feynman introduced new concepts in superfluidity and quantum
gravity. He was one of the recipients of the Nobel Prize for Physics
in 1965.
Feynman’s talk at MIT was entitled Simulating Physics with Computers.
In it, he argued that conventional computers will never be up to the
job of accurately simulating the behaviour of subatomic particles. He
proposed that a new kind of computer was needed. A computer that
would use quantum effects to simulate physical systems. His idea was
that the weird behaviour of subatomic particles could be exploited to
perform calculations at incredibly high speeds. Feynman dubbed his
theoretical machine a quantum computer.
For more than a decade, Feynman’s idea was an intellectual
curiosity—something that mathematicians and physicists toyed with,
but no one took seriously. Actually, building such a machine would
be an inordinately complex exercise. Plus, there didn’t seem to be
much point in doing it. Conventional electronic computers were good
enough for most tasks.
Then, in 1994, Peter Shor, a Professor of Applied Mathematics at MIT,
changed mainstream thinking on quantum computers. He unveiled an
algorithm that could perform fast prime factorization on a quantum
computer. If a quantum computer could be built, Shor’s algorithm
would be orders of magnitude faster than any previous method.
Conventional computers represent information by means of voltage
levels on microscopic wires. If the voltage level on a wire is high, then
the wire represents a one. In contrast, a low voltage level indicates a
zero value. This two-level system is called binary since every wire can
only have one of two values—0 or 1 (see Chapter 7). Crucially, at any
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Quantum Computers 225

moment in time, the voltage level on each wire has a single value and
so represents a single binary digit (or bit). Therefore, calculations have
to be performed one after another.
In contrast, quantum computers represent information using the
properties of subatomic, or quantum, particles. Various physical prop-
erties of subparticles can be used. One option is the spin of an electron.
An upward spin might represent a 1, downward a 0. The big advantage
of using the properties of subatomic particles is that, in the quantum
world, particles can exist in multiple states at the same time. This
strange behaviour is encapsulated in the principle of superposition. The effect
was uncovered by physicists in the early part of the twentieth century.
An electron can spin with all possible orientations simultaneously.
Exploiting this effect to represent data means that a single electron can
represent 0 and 1 simultaneously. This phenomena gives rise to the
basic unit of information in a quantum computer—the quantum bit,
or qubit.
Quantum computers become exponentially more powerful as
qubits are added. A single qubit can represent two values—0 and
1—simultaneously. Two qubits allow four values—00, 01, 10, and 11—
to be represented at the same time. A ten-qubit system can capture
all decimal values from 0 to 1,023, inclusive, simultaneously. When a
quantum computer performs an operation, it is applied to all states
at the same time. For example, adding one to a ten-qubit system
performs 1,024 additions at once. On a conventional computer, these
1,024 additions would have to be performed one after another. This
effect bestows on quantum computers the potential for exponential
acceleration in calculation.
There is a snag, though. Measuring the value of a qubit collapses its
state. This means that the qubit settles to a single value when its physical
state is measured. Thus, even though a ten-qubit system can perform
1,024 additions simultaneously, only one result can be retrieved. Worse
again, the output retrieved is selected at random from the 1,024 pos-
sibilities. The collapsed result of the addition of 1 could be any value
from 1 to 1,024. Clearly, random selection of the output is not desirable.
Mostly, we want to input some data and recover a particular result. The
solution to this problem is an effect known as interference. It is sometimes
possible to force unwanted states to destructively interfere with each
another. In this way, the unwanted results can be removed, leaving the
single, desired outcome behind.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

226 Next Steps

Quantum computers are well suited to puzzles that require many

alternatives to be evaluated and a single result returned. Combinato-
rial optimization problems match this requirement rather nicely (see
Chapter 6). The Travelling Salesman Problem, for example, requires
that the length of all possible city tours be evaluated and only the
shortest returned. This fits perfectly with the architecture of a quantum
computer, provided that the suboptimal solutions can be made to in-
terfere. For combinatorial optimization problems, quantum computers
promise performance far in excess of the world’s fastest supercomput-
ers. A working quantum computer would revolutionize challenging
problems such as drug discovery, materials design, and scheduling. It
would also crack the prime factorization problem.
Shor’s algorithm for finding prime factors is slow on a conventional
computer but is well-suited to a quantum computer. The algorithm
begins by guessing one of the primes. Of course, this guess is almost
certainly incorrect. Rather than guess again, Shor’s algorithm seeks
to improve this guess. It does this by multiplying the guess by itself
over and over again. Every time, it divides the original large number
by the result of the multiplication and stores the remainder. After a
large number of repetitions, the sequence of remainder values displays
a pattern—the series repeats with a fixed period (see clock arithmetic
in Chapter 7). Shor’s algorithm determines this period by means of the
Fourier transform (see Chapter 2). The peak of the Fourier transform
output identifies the sequence’s period. A multiple of the sought-after
prime can be calculated as the original guess to the power of the period
divided by two, all minus one.
At this point, the algorithm has the original large number and a
multiple of one of its prime factors. These two numbers are both
multiples of the desired prime. Finding the greatest common divisor
of two numbers is relatively straight-forward. Euclid’s algorithm can
be applied (see Chapter 1). Euclid’s algorithm repeatedly subtracts one
number from the other until the working values are equal. When this
happens, the working values are equal to the greatest common divisor.
In Shor’s algorithm, the greatest common divisor is one of the prime
factors. The other prime factor can be found simply by dividing the
original large number by this prime.
The procedure doesn’t work every time. Success depends on the
initial guess. If the routine fails, the steps are repeated starting with
a different guess. In ninety-nine per cent of cases, Shor’s algorithm
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Quantum Computers 227

produces the prime factors in ten or fewer iterations (see the Appendix
for more details).
On a conventional computer, multiplying the guess by itself over and
over again is very slow. The loop has to be repeated a large number of
times before any pattern appears. On a quantum computer, these mul-
tiplications can be performed simultaneously, thanks to superposition.
After that, a quantum Fourier transform can be used to cancel all but
the strongest repeating pattern. This gives the period of the remainder
sequence, which can be collapsed and measured. Euclid’s algorithm
is then performed on a conventional computer. Superposition and
interference allow the quantum computer to perform Shor’s algorithm
amazingly quickly.
Teams at Google, IBM, Microsoft, and a handful of start-ups are now
chasing the quantum computing dream. Their devices bear greater
resemblance to large physics experiments than supercomputers. Build-
ing a quantum computer requires design and subatomic fabrication
of quantum logic gates. Measuring and controlling the state of the
subatomic particles requires incredibly precise equipment. To perform
reliable measurements, the qubits must be cooled to near absolute zero
(–273°C).
To date, computation with up to seventy-two qubits has been
demonstrated. In theory, seventy-two qubits should provide immense
computing power. However, in practice, quantum noise affects perfor-
mance. Minute fluctuations in the state of the subatomic particles can
lead to errors in computation. Teams compensate for this by dedicating
some of the qubits to error correction (see Chapter 7). The downside
is that fewer qubits are available for computation. On the face of it,
the solution appears to be straightforward—simply add more qubits.
However, there is a worry. What if more qubits just mean more noise
and errors? What if none are available for computation?
In October 2019, a team from Google claimed that their quantum
computer had attained quantum supremacy. The group stated that the
computer had performed a computation that could not conceivably
be completed on a conventional computer. The program checked that
the output of a quantum random number generator was truly random.
Their Sycamore quantum computing chip completed the task in 200
seconds using fifty-three qubits. The team estimated that the same
calculation would take more than 10,000 years on a supercomputer.
IBM begged to differ. They calculated that the task could be performed
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

228 Next Steps

in two and a half days on a supercomputer. Not quantum supremacy

then, but, still, the difference between 4 and 3,600 minutes is stark.
A great many challenges remain. However, it does appear that the
quantum computer designers are on to something big.

Not The End

Algorithms have come a long way since they were first etched into clay
tablets in ancient Mesopotamia. The first computers transformed their
importance and capability. Since the invention of the integrated circuit,
the power of algorithms has risen exponentially. It may be that further
acceleration will come from quantum computing. It is difficult to see
more than a few years ahead, but it seems that AI appears set to radically
change the way that our world works.
A dearth of translators has meant that thousands of clay tablets from
ancient Mesopotamia lie unread in the museums of the world. Today,
the latest AI algorithms are being put to the task of autonomously
translating 67,000 administrative tablets from twenty-first century bce
southern Mesopotamia. In perhaps the longest of full circles, the newest
algorithms are about to interpret the oldest.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Appendix

PageRank Algorithm
Take the table of link counts as input.
Calculate the PageRanks as the number of incoming links for a page
divided by the average number of incoming links.
Repeat the following:
Repeat the following for every column:
Set a running total to zero.
Repeat the following for every entry in the column:
Look up the current PageRank for the row.
Multiply by the number of links between the row and
column.
Divide by the total number of outgoing links for the row.
Multiply by the damping factor.
Add to the running total.
Stop repeating when all entries in the column have been
processed.
Add the damping term to the running total.
Store this value as the new PageRank for the column.
Stop repeating when all columns have been processed.
Stop repeating when the change in the PageRanks is small.
Output the PageRanks.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

230 Appendix

Artificial Neural Network Training

Take the training dataset and network topology as input.
Populate the topology with random parameters.
Repeat the following:
Repeat for every training example:
Apply the input to the network.
Calculate the network output using forward propagation.
Calculate the error between the actual and desired outputs.
Repeat for every layer moving backwards through the
network:
Repeat for every neuron in the layer:
Repeat for every weight and bias in the neuron:
Determine the relationship between the
parameter and the error.
Calculate a correction value for the parameter.
Multiply the correction by the training rate.
Subtract this value from the parameter.
Stop repeating when the neuron has been updated.
Stop repeating when the layer has been updated.
Stop repeating when network has been updated.
Stop repeating when the training dataset has been exhausted.
Stop repeating when there is no further reduction in the error.
Freeze the parameters.
Training is complete.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

Appendix 231

Bitcoin Algorithm
The bitcoin sender:
Creates a transaction recording the sender’s public key, the
receiver’s public key, the amount, and the IDs of the inputs to
the transaction.
Appends a digital signature to the transaction.
Broadcasts the signed transaction to the Bitcoin network.
The computers on the Bitcoin network:
Check that the signature is authentic.
Check that the input transactions have not been spent.
Incorporate the transaction in a candidate block.
Link the candidate to the chain.
Repeat the following steps:
Generate a random number and append it to the block.
Calculate the hash for the block.
Stop repeating when the hash is less than the threshold or
abandon the hunt when another computer wins the race.
Broadcast the valid block to the network.
The bitcoin receiver:
Accepts the transaction when it and five more blocks have been
added to the chain.
 OUP CORRECTED PROOF – FINAL, 12/7/2020, SPi

232 Appendix

Shor’s Algorithm
Take a large number as input.
Repeat the following steps:
Take a prime number as a guess.
Store the guess in memory.
Create an empty list.
Repeat the following steps:
Multiplying the value in memory by the guess.
Update the value in memory.
Calculate the remainder after dividing the input by the value
in memory.
Append this remainder to the list.
Stop repeating after a large number of repetitions.
Apply the Fourier transform to the list of remainders.
Identify the period of the strongest harmonic.
Calculate the guess to the power of the period divided by two, all
minus one.
Apply Euclid’s algorithm to this value and the input.
Stop repeating when the value returned is a prime factor of the input.
Divide the input by the prime factor.
Output both prime factors.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Notes

Introduction
(Page 1) Strictly speaking, even addition is an algorithm.
(Page 1) A common misconception is that the word ‘algorithm’ is synonymous
with ‘method’. The two words are not equivalent. A method is a series of steps.
An algorithm is a series of steps that solve an information problem.
(Page 3) In this example, the books are considered as symbols representing the
titles of the books. Rearranging the books—the symbols—has the effect of
sorting the titles.

Chapter 1 Ancient Algorithms

(Page 10) The Incas were the only Bronze age civilization that, as far as we know,
did not invent writing. Egyptian mathematics was recorded on papyrus. As re-
sult, it is suspected that much has been lost. Ancient Egyptian mathematics was
practical in nature and revolved around numerical calculation. Mesopotamian
mathematics was more explicit in the usage, application, and description of
algorithms.
(Page 11) Originally a city state in central Mesopotamia, the Akkadian Empire
grew to encompass the bulk of the land between the rivers and parts of the
Levant.
(Page 16) Heron’s algorithm is a simplification of the more general Newton–
Raphson method.
(Page 16) The approximation algorithm can be accelerated by dividing two by
the most recent approximation, rather than taking the value halfway between
the most recent approximations.
(Page 17) In 1994, Jerry Bonnell and Robert Nemiroff programmed a VAX
computer to enumerate the square root of two to ten million digits. Bonnell
and Nemiroff did not disclose which algorithm they used.
(Page 20) The original version of Euclid’s algorithm used subtraction. It is
possible to use division instead. In some cases, using division is faster, but it must
be remembered that a single division operation is, in fact, a series of subtraction
operations. Alternatively, division can be performed as subtraction in the log
domain.
Chapter 2 Ever-Expanding Circles
(Page 27) It has been claimed that Archimedes’ screw was invented by the
Babylonians and the design transported to Egypt in the time of Archimedes.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

234 Notes

(Page 28) Archimedes did not have the benefit of the sine, cosine, and tangent
trigonometric functions that we use today. The length of a side of the inner
hexagon is 2r sin( π6 ). The angle is the angle from the centre to the bisection of
the side. The length of a side of the outer hexagon is 2r tan( π6 ).
(Page 29) Archimedes’ algorithm was finally supplanted by calculations based
on infinite series.
(Page 30) Quadratic algorithms are of the form ax2 + bx + c = 0 where a, b,
and c are known constants, or coefficients, and x is the unknown value to be
determined.
(Page 30) The Compendious Book on Calculation by Completion and Balancing was translated
into Latin by Robert of Chester around 1145.
(Page 31) Other cultures developed decimal numbers systems, including the
Chinese and Egyptians. However, they used different numerals (digit represen-
tations) and, by and large, used alternative positional systems.
(Page 33) More precisely, Fourier claimed that any function of a variable could
be expressed as the summation of a series of sinusoidal functions whose periods
are power of two divisors of the period of the original function. The Fourier
series had been previously used by Leonhard Euler, Joseph Louis Lagrange, and
Carl Friedrich Gauss. However, Fourier’s work served to popularize the concept
and was the basis of later work.
(Page 35) In the Fourier transform example, I omit the DC (constant) compo-
nent for simplicity.
(Page 37) Tukey also has the distinction of coining the terms ‘software’ and ‘bit’.

Chapter 3 Computer Dreams

(Page 42) The finished section of the Difference Engine No. 1 is now on display
in the Science Museum in London. The unit is cuboid—just over 60 cm high,
60 cm wide and almost 45 cm in length. A wooden base supports a metal
framework containing three stacks of brass discs. The discs are labelled with
the decimal digits and are interconnected by an intricate mechanism of shafts,
levers and gears. A crank and a series of cogwheels sit atop the engine, above
a metal plate. The quality and precision of Clement’s workmanship is evident.
However, to the modern eye, the device looks more like an ingenious, Victorian
cash register than the makings of a computer.
(Page 45) Lovelace’s participation in the Analytic Engine project has been
overstated in some quarters. She was not involved in the design of the machine
itself. However, she did understand what it did, how it could be used for
computation, and how to program it. She challenged Babbage and pushed him
to explain his methods. Not only did she consider what the machine was, but
also she envisioned what it might become. Perhaps her greatest achievement
lay in communicating Babbage’s extraordinary ideas to a wider audience.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Notes 235

(Page 45) A section of the Analytic Engine was assembled and is now in the
Science Museum, London.
(Page 45) Macabrely, half of Babbage’s brain is on display in the Science Museum,
London. The other half is held in the Royal College of Surgeons. Menabrea,
author of the original paper on the Analytic Engine, went on to become Prime
Minister of Italy (1867–1869).
(Page 50) Turing’s supervisor at Princeton, Alonzo Church, came up with an
alternative, calculus-based proof at roughly the same time. Turing’s proposal
was closely related to earlier work by Kurt Gödel.
(Page 52) Turing’s original description of the Turing Test, rather oddly, equates
differentiating between a computer and human with differentiating between
a man and woman. One wonders if there was a subtext regarding his own
homosexuality.
(Page 53) It has been reported that the Apple logo was emblematic of the apple
found by Turing’s bedside. When asked, Steve Jobs replied that it wasn’t, but he
wished it had been.
(Page 54) A hack to make the Z3 Turing Complete was published in 1998.
(Page 54) Under the direction of George Stibitz, Bell Labs also developed a relay-
based calculator.

Chapter 4 Weather Forecasts

(Page 59) Herein, I use the word ‘computer’ as shorthand for ‘pseudo Tur-
ing Complete computer’. Pseudo in that they do not have infinite memory.
Pseudo Turing Complete computers are digital in that they process numbers
and these numbers represent information. So-called ‘analog computers’ are
fixed-function devices that use a continuous physical quantity to represent
information.
(Page 63) A key part of Larson’s ruling was ‘Eckert and Mauchly did not
themselves first invent the automatic electronic digital computer, but instead
derived that subject matter from one Dr. John Vincent Atanasoff.’
(Page 66) The law of large numbers states that the average outcome of a number
of trials of a random process tends to the true value as more trials as performed.
(Page 67) Enrico Fermi previously experimented with a version of the Monte
Carlo method but did not publish on the topic.
(Page 67) Metropolis later lived up to his box office name by appearing as a
scientist in a Woody Allen movie.
(Page 71) Henri Poincaré identified a chaotic system in the 1880s in the form
of three bodies orbiting one another. He also developed theory to investigate
the effect.
(Page 73) Today, the number of transistors in an integrated circuit doubles every
twenty-four months—a slight deceleration.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

236 Notes

Chapter 5 Artificial Intelligence Emerges

(Page 77) The British Broadcasting Corporation recorded three of Strachey’s
computer compositions: The National Anthem, Baa Baa Black Sheep, and In The Mood.
Restorations of the historic recordings are now online.
(Page 86) In 1971, Shaw left RAND to work as a software and programming
consultant. He passed away in 1991.
(Page 90) Chinook, a Checkers-playing program written by Jonathan Schaeffer,
defeated world champion Marion Tinsley in 1994.
(Page 91) Samuel’s learning and minimax procedures drew on suggestions
made by Claude Shannon in a 1950 paper regarding Chess. Unlike Samuel,
Shannon did not develop an actual program.
(Page 91) Newell and Simon did make three other predictions that did
come true.

Chapter 6 Needles in Haystacks

(Page 96) Currently, the fastest algorithm for solving the Travelling Salesman
Problem has exponential complexity. In 1976, Nicos Christofides came up with
an algorithm that quickly produces routes that are guaranteed to be at most
fifty percent worse than the minimum route. Since then, fast approximation
algorithms have been improved so as to give a guarantee of forty percent of the
minimum.
(Page 98) In the worst case, Quicksort takes as many operations as
Insertion Sort.
(Page 102) The only Millennium Problem to have been solved so far is the
Poincaré Conjecture by Grigori Perelman in 2003.
(Page 103) George Forsythe has been credited with coining the term ‘computer
science’ in a paper published in 1961, but the term is older than that. Luis Fein
used it in a paper in 1959 to describe university computing schools.
(Page 110) The original NRMP algorithm was developed by John Mullin and
J.M. Stalnaker, prior to the adoption of the Boston Pool algorithm. The Gale–
Shapley algorithm was rejected twice for being too simple before it was finally
published in 1962.
(Page 110) Prior to Holland’s work, Nils Barricelli and Alexander Fraser
used computer algorithms to model and study biological evolutionary
processes. However, their proposals lacked certain key elements contained in
Holland’s work.
(Page 111) It has been claimed that Holland was the first person to receive a PhD
in Computer Science in the US. In fact, he was enrolled to a Communication
Sciences graduate programme at the University of Michigan, not a Computer
Science programme. The first two PhD degrees in Computer Science in the
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Notes 237

US were awarded on the same day—7 June 1965. The recipients were Sister
May Kemmar at the University of Wisconsin and Irving Tang at Washington
University in St. Louis.
(Page 111) Fisher dedicated his book to Darwin’s son, Leonard Darwin, with
whom Fisher had a long friendship and who provided much support in the
writing of the book.
(Page 113) Paradoxically, Holland used the success of natural evolution to
justify his work on genetic algorithms, whereas biologists employed Holland’s
algorithms to support their arguments for the existence of natural evolution.

Chapter 7 The Internet

(Page 120) There is some disagreement about the role of Leonard Kleinrock
in the development of packet-switching. In my view, during his PhD at MIT,
Kleinrock developed mathematical analyses that were applicable to packet-
switched networks, but he did not invent packet-switching.
(Page 126) The word ‘Internet’, a contraction of internetworking, seems to have
been coined by Vint Cerf and two colleagues at Stanford—Yogen Dalal and Carl
Sunshine.
(Page 129) The ISBN check digit is calculated from the other twelve digits by
multiplying by 1 or 3 alternating, adding the results together, extracting the
last decimal digit from the sum, subtracting the digit from 10, and, if necessary,
replacing a resulting 10 with a 0. For example, the novel A Game of Thrones has the
ISBN 978-000754823-1. The checksum is (9 × 1) + (3 × 7) + (8 × 1) + (3 × 0) +
(0×1)+(3×0)+(7×1)+(3×5)+(4×1)+(3×8)+(2×1)+(3×3) = 99. The
check digit is then 10 − 9 = 1. There is a nine in ten chance that a single digit
transcription error will be detected by cross-validation of the ISBN check digit.
(Page 132) The quick way to determine the bit position in-error is to write the
parity check results down in reverse order. In the example, this gives 0011 where
0 indicates an even count (no error in that group) and 1 indicates an odd count
(an error in that group). This value can be interpreted as a binary number 0011
giving the position of the error, in this case, position 3.
(Page 135) Merkle followed up with a paper outlining his own ideas in 1978.
(Page 136) The Alice, Bob, and Eve (eavesdropper) characters were invented by
Ron Rivest, Adi Shamir, and Leonard Adleman to explain their new encryption
algorithm. The characters have taken on a life of their own and are now
routinely referred to in papers on cryptography and security.
(Page 139) Formally, the totient is the number of integers less than a number
that are co-prime to it, i.e. they share no factors. The public exponent is a
number between one and the totient where the chosen number and the totient
are co-prime. Co-prime means that they must not be both evenly divisible by
the same number, other than one. A simple solution is just to choose the public
exponent as a prime number less than the totient.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

238 Notes

Chapter 8 Googling the Web

(Page 146) Mosaic was soon displaced by Netscape Navigator. Microsoft licensed
Mosaic for the development of Internet Explorer.

Chapter 9 Facebook and Friends

(Page 166) In truth, a lot of other factors can be used. For example, good
recommenders don’t just select similar users and similar movies. All users and
movies can be used as predictors. Imagine that Ken and Jill never agree on
movies. Let’s say Ken’s ratings are always the exact opposite of Jill’s. If Ken says
1 star, then Jill says 5 stars, and so on. Even though their scores are dissimilar,
Ken’s ratings are, in fact, the perfect predictor of Jill’s. Just subtract Ken’s score
from 6. The fact that their ratings histories are always dissimilar is helpful
information.

Chapter 10 America’s Favourite Quiz Show

(Page 172) I don’t discuss the Deep Blue versus Kasparov match in the book
because it is more a computer chip design story than an algorithm story.
(Page 172) Gary Tesauro of TD-Gammon fame worked on the game playing
strategy elements of the system.

Chapter 11 Mimicking the Brain

(Page 186) When counting the number of layers in a network, the input layer
is excluded.
(Page 188) Minsky and Rosenblatt both attended the Bronx High School of
Science.
(Page 188) I have seen claims of forthright face-to-face debates between Minsky
and Rosenblatt but I have not come across any first-hand accounts of same.
(Page 195) Hinton is a great-great-grandson of George Boole.
(Page 197) The term Deep Learning was coined by Rina Dechter in 1986 in
reference to machine learning and in 2000 by Igor Aizenberg in reference to
neural networks.
(Page 201) Google took over funding of the Turing Award in 2014, quadrupling
the prize fund to one million dollars.
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Notes 239

Chapter 13 Next Steps

(Page 217) The smallest Bitcoin unit is the satoshi—one hundred millionth of
a bitcoin.
(Page 217) The word for the Bitcoin concept is in title case, whereas references
to bitcoins are in lower case.
(Page 224) In the event that RSA is broken, Bitcoin can switch to postquantum
cryptographical techniques such as elliptic curve cryptography.
(Page 224) Feynman’s 1985 best-selling autobiography—Surely You’re Joking, Mr.
Feynman!—made him truly famous.
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Permissions

Algorithm definition: © Oxford University Press. Courtesy Oxford Uni-

versity Press.
Chapter 1 and 4 epigraphs: © Oxford University Press. Courtesy Oxford
University Press.
Cistern algorithm: Reprinting privileges granted by permission of the
Association for Computing Machinery.
Stanislaw Ulam: By Los Alamos National laboratory. Unless otherwise
indicated, this information has been authored by an employee or em-
ployees of the Los Alamos National Security, LLC (LANS), operator
of the Los Alamos National Laboratory under Contract No. DE-AC52-
06NA25396 with the US. Department of Energy. The US Government
has rights to use, reproduce, and distribute this information. The public
may copy and use this information without charge, provided that this
Notice and any statement of authorship are reproduced on all copies.
Neither the Government nor LANS makes any warranty, express or
implied, or assumes any liability or responsibility for the use of this
information.
Chapter 8 epigraph: This excerpt was originally published in The Atlantic
and is republished here with The Atlantic’s permission.
Chapter 10 epigraph: Courtesy Ken Jennings.
Chapter 12 epigraph: © Robert W. Foster. Courtesy Robert W. Foster.
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography

1. Hoare, C.A.R., 1962. Quicksort. The Computer Journal, 5(1), pp. 10–16.
2. Dalley, S., 1989. Myths from Mesopotamia. Oxford: Oxford University Press.
3. Finkel, I., 2014. The Ark before Noah. Hachette.
4. Rawlinson, H.C., 1846. The Persian cuneiform inscription at Behistun,
decyphered and translated. Journal of the Royal Asiatic Society of Great Britain
and Ireland, 10, pp. i–349.
5. Knuth, D.E., 1972. Ancient Babylonian algorithms. Communications of the
ACM, 15(7), pp. 671–7.
6. Fowler, D. and Robson, E., 1998. Square root approximations in old
Babylonian mathematics: YBC 7289 in context. Historia Mathematica, 25(4),
pp. 366–78.
7. Fee, G.J., 1996. The square root of 2 to 10 million digits. http://www.plouffe.
fr/simon/constants/sqrt2.txt. (Accessed 5 July 2019).
8. Harper, R.F., 1904. The Code of Hammurabi, King of Babylon. Chicago: The
University of Chicago Press.
9. Jaynes, J., 1976. The Origin of Consciousness in the Breakdown of the Bicameral Mind.
New York: Houghton Mifflin Harcourt.
10. Boyer, C.B. and Merzbach. U.C., 2011. A History of Mathematics. Oxford: John
Wiley & Sons.
11. Davis, W.S., 1913. Readings in Ancient History, Illustrative Extracts from the Source:
Greece and the East. New York: Allyn and Bacon.
12. Beckmann, P., 1971. A history of Pi. Boulder, CO: The Golem Press.
13. Mackay, J.S., 1884. Mnemonics for π , π1 , e. Proceedings of the Edinburgh Mathe-
matical Society, 3, pp. 103–7.
14. Dietrich, L., Dietrich, O., and Notroff, J., 2017. Cult as a driving force of
human history. Expedition Magazine, 59(3), pp. 10–25.
15. Katz, V., 2008. A History of Mathematics. London: Pearson.
16. Katz, V. J. ed., 2007. The Mathematics of Egypt, Mesopotamia, China, India, and Islam.
Princeton, NJ: Princeton University Press.
17. Palmer, J., 2010. Pi record smashed as team finds two-quadrillionth
digit – BBC News [online]. https://www.bbc.com/news/technology-
11313194, September 16 2010. (Accessed 6 January 2019).
18. The Editors of Encyclopaedia Britannica, 2020. Al-Khwārizmī. In Ency-
clopaedia Britannica [online]. https://www.britannica.com/biography/al-
Khwarizmi (Accessed 20 May 2020).
19. LeVeque, W.J. and Smith, D.E., 2019. Numerals and numeral systems. In
Encyclopaedia Britannica [online]. https://www.britannica.com/science/
numeral. (Accessed 19 May 2020).
20. The Editors of Encyclopaedia Britannica, 2020. French revolution. In Ency-
clopaedia Britannica [online]. https://www.britannica.com/event/French-
Revolution. (Accessed 19 May 2020).
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

244 Bibliography

21. Cooley, J.W. and Tukey, J.W., 1965. An algorithm for the machine
calculation of complex Fourier series. Mathematics of Computation, 19(90),
pp. 297–301.
22. Rockmore, D.N., 2000. The FFT: An algorithm the whole family can use.
Computing in Science & Engineering, 2(1), pp. 60–4.
23. Anonymous, 2016. James William Cooley. New York Times.
24. Heidenman, C., Johnson, D., and Burrus, C., 1984. Gauss and the history
of the fast Fourier transform. IEEE ASSP Magazine, 1(4), pp. 14–21.
25. Huxley, T.H., 1887. The Advance of Science in the Last Half-Century. New York:
Appleton and Company.
26. Swade, D., 2000. The Cogwheel Brain. London: Little, Brown.
27. Babbage, C., 2011. Passages from the Life of a Philosopher. Cambridge: Cambridge
University Press.
28. Menabrea, L.F. and King, A., Countess of Lovelace, 1843. Sketch of
the analytical engine invented by Charles Babbage. Scientific Memoirs, 3,
pp. 666–731.
29. Essinger, J., 2014. Ada’s algorithm: How Lord Byron’s daughter Ada Lovelace Launched
the Digital Age. London: Melville House.
30. Kim, E.E. and Toole, B.A., 1999. Ada and the first computer. Scientific
American, 280(5), pp. 76–81.
31. Isaacson, W., 2014. The Innovators. New York: Simon and Schuster.
32. Turing, S., 1959. Alan M. Turing. Cambridge: W. Heffer & Sons, Ltd.
33. Turing, A.M., 1937. On computable numbers, with an application to the
Entscheidungsproblem. Proceedings of the London Mathematical Society, s2–42(1),
pp. 230–65.
34. Davis, M., 1983. Computability and Unsolvability. Mineola, NY: Dover Publica-
tions.
35. Strachey, C., 1965. An impossible program. The Computer Journal, 7(4), p. 313.
36. Turing, A.M., 1950. Computing machinery and intelligence. Mind, 59(236),
pp. 433–60.
37. Copeland, B. Jack., 2014. Turing. Oxford: Oxford University Press.
38. Abbe, C., 1901. The physical basis of long-range weather forecasts. Monthly
Weather Review, 29(12), pp. 551–61.
39. Lynch, P., 2008. The origins of computer weather prediction and climate
modeling. Journal of Computational Physics, 227(7), pp. 3431–44.
40. Hunt, J.C.R., 1998. Lewis Fry Richardson and his contributions to mathe-
matics, meteorology, and models of conflict. Annual Review of Fluid Mechanics,
30(1), pp. xiii–xxxvi.
41. Mauchly, J.W., 1982. The use of high speed vacuum tube devices for cal-
culating. In: B. Randall, ed., The Origins of Digital Computers. Berlin: Springer,
pp. 329–33.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 245

42. Fritz, W.B., 1996. The women of ENIAC. IEEE Annals of the History of Comput-
ing, 18(3), pp. 13–28.
43. Ulam, S., 1958. John von Neumann 1903–1957. Bulletin of the American Math-
ematical Society, 64(3), pp. 1–49.
44. Poundstone, W., 1992. Prisoner’s Dilemma. New York: Doubleday.
45. McCorduck, P., 2004. Machines Who Think. Natick, MA: AK Peters.
46. Goldstein, H.H., 1980. The Computer from Pascal to von Neumann. Princeton, NJ:
Princeton University Press.
47. Stern, N., 1977. An Interview with J. Presper Eckert. Charles Babbage Institute,
University of Minnesota.
48. Von Neumann, J., 1993. First draft of a report on the EDVAC. IEEE Annals
of the History of Computing, 15(4), pp. 27–75.
49. Augarten, S., 1984. A. W. Burks, ‘Who invented the general-purpose
electronic computer?’ In Bit by bit: An Illustrated History of Computers. New
York: Ticknor & Fields. Epigraph, Ch. 4.
50. Kleiman, K., 2014. The computers: The remarkable story of the
ENIAC programmers. Vimeo [online]. https://vimeo.com/ondemand/
eniac6. (Accessed 11 March 2019).
51. Martin, C.D., 1995. ENIAC: Press conference that shook the world. IEEE
Technology and Society Magazine, 14(4), pp. 3–10.
52. Nicholas Metropolis. The beginning of the Monte Carlo method. Los
Alamos Science, 15(584), pp. 125–30.
53. Eckhardt, R., 1987. Stan Ulam, John von Neumann, and the Monte Carlo
method. Los Alamos Science, 15(131–136), p. 30.
54. Wolter, J., 2013. Experimental analysis of Canfield solitaire. http://politaire.
com/article/canfield.html. (Accessed 20 May 2020).
55. Metropolis, N. and Ulam, S., 1949. The Monte Carlo method. Journal of the
American Statistical Association, 44(247), pp. 335–341.
56. Charney, J.G. and Eliassen, A., 1949. A numerical method for predicting
the perturbations of the middle latitude westerlies. Tellus, 1(2), pp. 38–54.
57. Charney, J.G., 1949. On a physical basis for numerical prediction of large-
scale motions in the atmosphere. Journal of Meteorology, 6(6), pp. 372–85.
58. Platzman, G.W., 1979. The ENIAC computations of 1950: Gateway to
numerical weather prediction. Bulletin of the American Meteorological Society,
60(4), pp. 302–12.
59. Charney, J.G., Fjörtoft, R., and von Neumann, J., 1950. Numerical inte-
gration of the barotropic vorticity equation. Tellus, 2(4), pp. 237–54.
60. Blair, C., 1957. Passing of a great mind. Life Magazine, 42(8), pp. 89–104.
61. Lorenz, E.N., 1995. The Essence of Chaos. Seattle: University of Washington
Press.
62. Lorenz, E.N., 1963. Deterministic nonperiodic flow. Journal of the Atmospheric
Sciences, 20(2), pp. 130–41.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

246 Bibliography

63. Epstein, E.S., 1969. Stochastic dynamic prediction. Tellus, 21(6), pp. 739–59.
64. European Centre for Medium-Range Weather Forecasts, 2020. Advancing
global NWP through international collaboration. http://www.ecmwf.int.
(Accessed 19 May 2020).
65. Lynch, P. and Lynch, O., 2008. Forecasts by PHONIAC. Weather, 63(11),
pp. 324–6.
66. Shannon, C.E., 1950. Programming a computer for playing chess. Philosoph-
ical Magazine, 41(314), pp. 256–75.
67. National Physical Laboratory, 2012. Piloting Computing: Alan Turing’s
Automatic Computing Engine. YouTube [online]. https://www.youtube.
com/watch?v=cEQ6cnwaY_s. (Accessed 27 October 2019).
68. Campbell-Kelly, M., 1985. Christopher Strachey, 1916–1975: A biographi-
cal note. Annals of the History of Computing, 7(1), pp. 19–42.
69. Copeland, J. and Long, J., 2016. Restoring the first recording of com-
puter music. https://blogs.bl.uk/sound-and-vision/2016/09/restoring-the-
first-recording-of-computer-music.html#. (Accessed 15 February 2019).
70. Foy, N., 1974. The word games of the night bird (interview with Christo-
pher Strachey). Computing Europe, 15, pp. 10–11.
71. Roberts, S., 2017. Christopher Strachey’s nineteen-fifties love machine.
The New Yorker, February 14.
72. Strachey, C., 1954. The ‘thinking’ machine. Encounter, III, October.
73. Strachey, C.S., 1952. Logical or non-mathematical programmes. In Proceed-
ings of the 1952 ACM National Meeting New York: ACM. pp. 46–9.
74. McCarthy, J., Minsky, M.L., Rochester, N., and Shannon, C.E., 2006. A
proposal for the Dartmouth summer research project on artificial intelli-
gence, August 31, 1955. AI Magazine, 27(4), pp. 12–14.
75. Newell, A. and Simon, H., 1956. The logic theory machine: A complex
information processing system. IRE Transactions on Information Theory. 2(3),
pp. 61–79.
76. Newell, A., Shaw, J.C., and Simon, H.A., 1959. Report on a general prob-
lem solving program. In Proceedings of the International Conference on Information
Processing. Paris: UNESCO. pp. 256–64.
77. Newell, A. and Simon, H., 1972. Human Problem Solving. New York: Prentice-
Hall.
78. Schaeffer, J., 2008. One Jump Ahead: Computer Perfection at Checkers. New York:
Springer.
79. Samuel, A.L., 1959. Some studies in machine learning using the game of
checkers. IBM Journal of Research and Development, 3(3), pp. 210–29.
80. McCarthy, J. and Feigenbaum, E.A., 1990. In memoriam: Arthur Samuel:
Pioneer in machine learning. AI Magazine, 11(3), p. 10.
81. Samuel, A.L., 1967. Some studies in machine learning using the game of
checkers. ii. IBM Journal of Research and Development, 11(6), pp. 601–17.
82. Madrigal, A.C., 2017. How checkers was solved. The Atlantic. July 19.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 247

83. Simon, H.A., 1998. Allen Newell: 1927–1992. IEEE Annals of the History of
Computing, 20(2), pp. 63–76.
84. CBS, 1961. The thinking machine. YouTube [online]. https://youtu.be/
aygSMgK3BEM. (Accessed 19 May 2020).
85. Dreyfus, H.L., 2005. Overcoming the myth of the mental: How philoso-
phers can profit from the phenomenology of everyday expertise. In:
Proceedings and Addresses of the American Philosophical Association, 79(2), pp. 47–65.
86. Nilsson, N.J., 2009. The Quest for Artificial Intelligence. Cambridge: Cambridge
University Press.
87. Schrijver, A., 2005. On the history of combinatorial optimization (till
1960). In: K. Aardal, G.L. Nemhauser, R. Weismantel, eds., Discrete optimiza-
tion, vol. 12. Amsterdam: Elsevier. pp. 1–68.
88. Dantzig, G., Fulkerson, R., and Johnson, S., 1954. Solution of a large-scale
traveling-salesman problem. Journal of the Operations Research Society of America,
2(4), pp. 393–410.
89. Cook, W., n.d. Traveling salesman problem. http://www.math.uwaterloo.
ca/tsp/index.html. (Accessed 19 May 2020).
90. Cook, S.A., 1971. The complexity of theorem-proving procedures. In:
Proceedings of the 3rd annual ACM Symposium on Theory of Computing. New York:
ACM. pp. 151–8.
91. Karp, R., n.d. A personnal view of Computer Science at Berkeley.
https://www2.eecs.berkeley.edu/bears/CS_Anniversary/karp-talk.html.
(Accessed 15 February 2019).
92. Garey, M.R. and Johnson, D.S., 1979. Computers and Intractability. New York:
W. H. Freeman and Company.
93. Dijkstra, E.W., 1972. The humble programmer. Communications of the ACM,
15(10), pp. 859–66.
94. Dijkstra, E.W., 2001. Oral history interview with Edsger W. Dijkstra. Tech-
nical report, Charles Babbage Institute, August 2.
95. Dijkstra, E.W., 1959. A note on two problems in connexion with graphs.
Numerische mathematik, 1(1), pp. 269–71.
96. Darrach, B., 1970. Meet Shaky: The first electronic person. Life Magazine,
69(21):58B–68B.
97. Hart, P.E., Nilsson, N.J., and Raphael, B., 1968. A formal basis for the
heuristic determination of minimum cost paths. IEEE Transactions on Systems
Science and Cybernetics, 4(2), pp. 100–7.
98. Hitsch, G.J., Hortaçsu, A., and Ariely, D., 2010. Matching and sorting in
online dating. The American Economic Review, 100(1), pp. 130–63.
99. NRMP. National resident matching program. http://www.nrmp.org.
(Accessed 19 May 2020).
100. Roth, A.E., 2003. The origins, history, and design of the resident match.
Journal of the American Medical Association, 289(7), pp. 909–12.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

248 Bibliography

101. Roth, A.E., 1984. The evolution of the labor market for medical interns
and residents: A case study in game theory. The Journal of Political Economy,
92, pp. 991–1016.
102. Anonymous, 2012. Stable matching: Theory, evidence, and practical de-
sign. Technical report, The Royal Swedish Academy of Sciences.
103. Kelly, K., 1994. Out of Control. London: Fourth Estate.
104. Vasbinder, J.W., 2014. Aha... That is Interesting!: John H. Holland, 85 years young.
Singapore: World Scientific.
105. London, R.L., 2013. Who earned first computer science Ph.D.? Communica-
tions of the ACM : blog@CACM, January.
106. Scott, N.R., 1996. The early years through the 1960’s: Computing at the
CSE@50. Technical report, University of Michigan.
107. Fisher, R.A., 1999. The Genetical Theory of Natural Selection. Oxford: Oxford
University Press.
108. Holland, J.H., 1992. Adaptation in Natural and Artificial Systems. Cambridge, MA:
The MIT Press.
109. Holland, J.H., 1992. Genetic algorithms. Scientific American, 267(1),
pp. 66–73.
110. Dawkins, R., 1986. The Blind Watchmaker. New York: WW Norton & Com-
pany.
111. Lohn, J.D., Linden, D.S., Hornby, G.S., Kraus, W.F., 2004. Evolutionary
design of an X-band antenna for NASA’s space technology 5 mission. In:
Proceedings of the IEEE Antennas and Propagation Society Symposium 2004, volume 3.
Monterey, CA, 20–25 June, pp. 2313–16. New York: IEEE.
112. Grimes, W., 2015. John Henry Holland, who computerized evolution, dies
at 86. New York Times, August 19.
113. Licklider, J.C.R., 1960. Man-computer symbiosis. IRE Transactions on Human
Factors in Electronics, 1(1), pp. 4–11.
114. Waldrop, M.M., 2001. The Dream Machine. London: Viking Penguin.
115. Kita, C.I., 2003. JCR Licklider’s vision for the IPTO. IEEE Annals of the History
of Computing, 25(3), pp. 62–77.
116. Licklider, J.C.R., 1963. Memorandum for members and affiliates of the
intergalactic computer network. Technical report, Advanced Research
Projects Agency, April 23.
117. Licklider, J.C.R., 1965. Libraries of the Future. Cambridge, MA: The MIT Press.
118. Licklder, J.C.R. and Taylor, R.W., 1968. The computer as a communication
device. Science and Technology, 76(2), pp. 1–3.
119. Markoff, J., 1999. An Internet pioneer ponders the next revolution. The New
York Times, December 20.
120. Featherly, K., 2016. ARPANET. In Encyclopedia Brittanica [online].
https://www.britannica.com/topic/ARPANET. (Accessed 19 May 2020).
121. Leiner, B.M., Cerf, V.G., Clark, D.D., Kahn, R.E., Kleinrock., L., Lynch,
D.C., Postel, J., Robers, L.G., and Wolff, S., 2009. A brief history of the
Internet. ACM SIGCOMM Computer Communication Review, 39(5), pp. 22–31.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 249

122. Davies, D.W., 2001. An historical study of the beginnings of packet switch-
ing. The Computer Journal, 44(3), pp. 152–62.
123. Baran, P., 1964. On distributed communications networks. IEEE Transactions
on Communications Systems, 12(1), pp. 1–9.
124. McQuillan, J., Richer, I., and Rosen, E., 1980. The new routing algorithm
for the ARPANET. IEEE Transactions on Communications, 28(5), pp. 711–19.
125. McJones, P., 2008. Oral history of Robert (Bob) W. Taylor. Technical
report, Computer History Museum.
126. Metz, C., 2012. Bob Kahn, the bread truck, and the Internet’s first com-
munion. Wired, August 13.
127. Vint, C. and Kahn, R., 1974. A protocol for packet network interconnec-
tion. IEEE Transactions of Communications, 22(5), pp. 637–48.
128. Metz, C., 2007. How a bread truck invented the Internet. The Register
[online]. https://www.theregister.co.uk/2007/11/12/thirtieth_anniversary
_of_first_internet_connection/. (Accessed 19 May 2020).
129. Anonymous, 2019. Number of internet users worldwide from 2005 to
2018. Statista [online]. https://www.statista.com/statistics/273018/number-
of-internet-users-worldwide/. (Accessed 19 May 2020).
130. Lee, J., 1998. Richard Wesley Hamming: 1915–1998. IEEE Annals of the History
of Computing, 20(2), pp. 60–2.
131. Suetonius, G., 2009. Lives of the Caesars. Oxford: Oxford University Press.
132. Singh, S., 1999. The Code Book: The Secret History of Codes & Code-breaking.
London: Fourth Estate.
133. Diffie, W. and Hellman, M., 1976. New directions in cryptography. IEEE
Transactions on Information Theory, 22(6), pp. 644–54.
134. Rivest, R.L., Shamir, A., Adleman, L., 1978. A method for obtaining digital
signatures and public-key cryptosystems. Communications of the ACM, 21(2),
pp. 120–6.
135. Gardner, M., 1977. New kind of cipher that would take millions of years
to break. Scientific American. 237(August), pp. 120–4.
136. Atkins, D., Graff, M., Lenstra, A.K., Leyland, P.C., 1994. The magic words
are squeamish ossifrage. In: Proceedings of the 4th International Conference on the
Theory and Applications of Cryptology. Wollongong, Australia. November 28
November-1 December 1994. pp. 261–77. NY: Springer.
137. Levy, S., 1999. The open secret. Wired, 7(4).
138. Ellis, J.H., 1999. The history of non-secret encryption. Cryptologia, 23(3),
pp. 267–73.
139. Ellis, J.H., 1970. The possibility of non-secret encryption. In: British
Communications-Electronics Security Group (CESG) report. January.
140. Bush, V., 1945. As we may think. The Atlantic. 176(1), pp. 101–8.
141. Manufacturing Intellect, 2001. Jeff Bezos interview on starting Amazon.
YouTube [online]. https://youtu.be/p7FgXSoqfnI. (Accessed 19 May 2020).
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

250 Bibliography

142. Stone, B., 2014. The Everything Store: Jeff Bezos and the Age of Amazon. New York:
Corgi.
143. Christian, B. and Griffiths, T., 2016. Algorithms to Live By. New York:
Macmillan.
144. Linden, G., Smith, B., and York, J., 2003. Amazon.com recommendations.
IEEE Internet Computing, 7(1), pp. 76–80.
145. McCullough, B., 2015. Early Amazon engineer and co-developer of
the recommendation engine, Greg Linden. Internet History Podcast
[online]. http://www.internethistorypodcast.com/2015/04/early-amazon-
engineer-and-co-developer-of-the-recommendation-engine-greg-
linden/#tabpanel6. (Accessed 15 February 19).
146. MacKenzie, I., Meyer, C., and Noble, S., 2013. How retailers can keep up
with consumers. Technical report, McKinsey and Company, October.
147. Anonymous, 2016. Total number of websites. Internet Live Stats [on-
line] http://www.internetlivestats.com/total-number-of-websites/. (Ac-
cessed 15 February 19).
148. Vise, D.A., 2005. The Google Story. New York: Macmillian.
149. Battelle, J., 2005. The birth of Google. Wired, 13(8), p. 102.
150. Page, L., Brin, S., Motwani, R., and Winograd, T., 1999. The PageRank
citation ranking: Bringing order to the web. Technical Report 1999–66,
Stanford InfoLab, November.
151. Brin, S. and Page, L., 1998. The anatomy of a large-scale hypertextual web
search engine. Computer Networks and ISDN Systems, 30(1–7), pp. 107–117.
152. Jones, D., 2018. How PageRank really works: Understanding Google. Ma-
jestic [blog]. https://blog.majestic.com/company/understanding-googles-
algorithm-how-pagerank-works/, October 25. (Accessed 12 July 2019).
153. Willmott, D., 1999. The top 100 web sites. PC Magazine, February 9.
154. Krieger, L.M., 2005. Stanford earns $336 million off Google stock. The
Mercury News, December 1.
155. Hayes, A., 2019. Dotcom bubble definition. Investopedia [online]. https://
www.investopedia.com/terms/d/dotcom-bubble.asp, June 25. (Accessed
19 July 2019).
156. Smith, B. and Linden, G., 2017. Two decades of recommender systems at
Amazon.com. IEEE Internet Computing, 21(3), pp. 12–18.
157. Debter, L., 2019. Amazon surpasses Walmart as the world’s largest
retailer. Forbes [online]. https://www.forbes.com/sites/laurendebter/
2019/05/15/worlds-largest-retailers-2019-amazon-walmart-alibaba/#
20e4cf4d4171. (Accessed 18 July 2019).
158. Anonymous, 2019. Tim Berners-Lee net worth. The Richest [online].
https://www.therichest.com/celebnetworth/celebrity-business/tech-
millionaire/tim-berners-lee-net-worth/. (Accessed 22 July 2019).
159. Anonymous, 2016. Internet growth statistics. Internet World Stats [on-
line]. http://www.internetworldstats.com/emarketing.htm. (Accessed 19
May 2020).
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 251

160. Conan Doyle, A., 1890. The sign of four. Lippincott’s Monthly Magazine.
February.
161. Kirkpatrick, D., 2010. The Facebook Effect. New York: Simon and Schuster.
162. Grimland, G., 2009. Facebook founder’s roommate recounts creation of
Internet giant. Haaretz [online]. https://www.haaretz.com/1.5050614. (Ac-
cessed 23 July 2019).
163. Kaplan, K.A., 2003. Facemash creator survives ad board. The Harvard Crimson,
November.
164. Investors Archive, 2017. Billionaire Mark Zuckerberg: Creating Facebook
and startup advice. YouTube [online]. https://youtu.be/SSly3yJ8mKU.
165. Widman, J., 2011. Presenting EdgeRank: A guide to Facebook’s Newsfeed
algorithm. http://edgerank.net. (Accessed 19 May 2020).
166. Anonymous 2016. Number of monthly active facebook users world-
wide. Statista [online] https://www.statista.com/statistics/264810/number-
of-monthly-active-facebook-users-worldwide/. (Accessed 20 May 2020).
167. Keating, G., 2012. Netflixed. London: Penguin.
168. Netflix, 2016. Netflix prize. http://www.netflixprize.com/. (Accessed 19
May 2020).
169. Van Buskirk, E., 2009. How the Netflix prize was won. Wired, September 22.
170. Thompson, C., 2008. If you liked this, you’re sure to love that. The New York
Times. November 21.
171. Piotte, M. and Chabbert, M., 2009. The pragmatic theory solution to the
Netflix grand prize. Technical report, Netflix.
172. Koren, Y., 2009. The Bellkor solution to the Netflix grand prize. Netflix Prize
Documentation, 81:1–10.
173. Johnston, C., 2012. Netflix never used its $1 million algorithm due to
engineering costs. Wired, April 16.
174. Gomez-Uribe, C.A. and Hunt, N., 2015. The Netflix recommender system:
Algorithms, business value, and innovation. ACM Transactions on Management
Information Systems, 6(4), pp. 13:1–19.
175. Ginsberg, J., Mohebbi, M.H., Patel, R.S., Brammer, L., Smolinkski, M.S.,
and Brilliant, L., 2009. Detecting influenza epidemics using search engine
query data. Nature, 457(7232), pp. 1012–14.
176. Cook, S., Conrad, C., Fowlkes, A.L., Mohebbi, M.H., 2011. Assessing
Google flu trends performance in the United States during the 2009
influenza virus A (H1N1) pandemic. PLOS ONE, 6(8), e23610.
177. Butler, D., 2013. When Google got flu wrong. Nature, 494(7436),
pp. 155.
178. Lazer, D., Kennedy, R., King, G., Vespignani, A., 2014. The parable of
Google Flu: Traps in big data analysis. Science, 343(6176), pp. 1203–5.
179. Lazer, D. and Kennedy, R., 2015. What we can learn from the epic failure
of Google flu trends. Wired, October 1.
180. Zimmer, B., 2011. Is it time to welcome our new computer overlords? The
Atlantic, February 17.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

252 Bibliography

181. Markoff, J., 2011. Computer wins on jeopardy: Trivial, it’s not. New York
Times, February 16.
182. Gondek, D.C., Lally, A., Kalyanpur, A., Murdock, J.W., Duboue, P.A.,
Zhang, L., Pan, Y., Qui, Z.M., and Welty, C., 2012. A framework for
merging and ranking of answers in DeepQA. IBM Journal of Research and
Development, 56(3.4), pp. 14:1–12.
183. Best, J., 2013. IBM Watson: The inside story of how the Jeopardy-winning
supercomputer was born, and what it wants to do next. TechRepublic
[online]. http://www.techrepublic.com/article/ibm-watson-the-inside-
story-of-how-the-jeopardy-winning-supercomputer-was-born-and-
what-it-wants-to-do-next/. (Accessed 15 February 2019).
184. Ferrucci, D.A., 2012. Introduction to this is Watson. IBM Journal of Research
and Development, 56(3.4), pp. 1:1–15.
185. IBM Research, 2013. Watson and the Jeopardy! Challenge. YouTube [on-
line]. https://www.youtube.com/watch?v=P18EdAKuC1U. (Accessed 15
September 2019).
186. Lieberman, H., 2011. Watson on Jeopardy, part 3. MIT Technology Re-
view [online]. https://www.technologyreview.com/s/422763/watson-on-
jeopardy-part-3/. (Accessed 15 September 2019).
187. Gustin, S., 2011. Behind IBM’s plan to beat humans at their own game.
Wired, February 14.
188. Lally, A., Prager, M., McCord, M.C., Boguraev, B.K., Patwardhan, S., Fan,
J., Fodor, P., and Carroll J.C., 2012. Question analysis: How Watson reads a
clue. IBM Journal of Research and Development, 56(3.4), pp. 2:1–14.
189. Fan, J., Kalyanpur, A., Condek, D.C., and Ferrucci, D.A., 2012. Automatic
knowledge extraction from documents. IBM Journal of Research and Develop-
ment, 56(3.4), pp. 5:1–10.
190. Kolodner, J.L., 1978. Memory organization for natural language data-base
inquiry. Technical report, Yale University.
191. Kolodner, J.L., 1983. Maintaining organization in a dynamic long-term
memory. Cognitive Science, 7(4), pp. 243–80.
192. Kolodner, J.L., 1983. Reconstructive memory: A computer model. Cogni-
tive Science, 7(4), pp. 281–328.
193. Lohr, S., 2016. The promise of artificial intelligence unfolds in small steps.
The New York Times, February 29. (Accessed 19 May 2020).
194. James, W., 1890. The Principles of Psychology. NY: Holt.
195. Hebb, D.O., 1949. The Organization of Behavior. NY: Wiley.
196. McCulloch, W.S. and Pitts, W., 1943. A logical calculus of the ideas
immanent in nervous activity. The Bulletin of Mathematical Biophysics, 5(4),
pp. 115–33.
197. Gefter, A., 2015. The man who tried to redeem the world with logic.
Nautilus, February 21.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 253

198. Whitehead, A.N. and Russell, B., 1910–1913. Principia Mathematica. Cam-
bridge: Cambridge University Press.
199. Anderson, J.A. and Rosenfeld, E., 2000. Talking Nets. Cambridge, MA: The
MIT Press.
200. Conway, F. and Siegelman, J., 2006. Dark Hero of the Information Age: In Search
of Norbert Wiener The Father of Cybernetics. New York: Basic Books.
201. Thompson, C., 2005. Dark hero of the information age: The original
computer geek. The New York Times, March 20.
202. Farley, B.W.A.C. and Clark, W., 1954. Simulation of self-organizing systems
by digital computer. Transactions of the IRE Professional Group on Information
Theory, 4(4), pp. 76–84.
203. Rosenblatt, F., 1958. The Perceptron: A probabilistic model for informa-
tion storage and organization in the brain. Psychological Review, 65(6), pp. 386.
204. Rosenblatt, F., 1961. Principles of neurodynamics. perceptrons and the
theory of brain mechanisms. Technical report, DTIC Document.
205. Anonymous, 1958. New Navy device learns by doing. The New York Times,
July 8.
206. Minsky, M. and Papert, S., 1969. Perceptrons. Cambridge, MA: The MIT Press.
207. Minksy, M., 1952. A neural-analogue calculator based upon a probability
model of reinforcement. Technical report, Harvard University Psycholog-
ical Laboratories, Cambridge, Massachusetts.
208. Block, H.D., 1970. A review of Perceptrons: An introduction to computa-
tional geometry. Information and Control, 17(5), pp. 501–22.
209. Anonymous, 1971. Dr. Frank Rosenblatt dies at 43; taught neurobiology
at Cornell. The New York Times, July 13.
210. Olazaran, M., 1996. A sociological study of the official history of the
Perceptrons controversy. Social Studies of Science, 26(3), pp. 611–59.
211. Werbos, P.J., 1990. Backpropagation through time: What it does and how
to do it. Proceedings of the IEEE, 78(10), pp. 1550–60.
212. Werbos, P.J., 1974. Beyond regression: New tools for prediction and analysis in the
behavioral sciences. PhD. Harvard University.
213. Werbos, P.J., 1994. The Roots of Backpropagation, volume 1. Oxford: John Wiley
& Sons.
214. Werbos, P.J., 2006. Backwards differentiation in AD and neural nets: Past
links and new opportunities. In: H.M. Bücker, G. Corliss, P. Hovland,
U. Naumann, and B. Norris, eds., Automatic differentiation: Applications, theory,
and implementations. Berlin: Springer. pp. 15–34.
215. Parker, D.B., 1985. Learning-logic: Casting the cortex of the human brain
in silicon. Technical Report TR-47, MIT, Cambridge, MA.
216. Lecun, Y., 1985. Une procédure d’apprentissage pour réseau a seuil asym-
metrique (A learning scheme for asymmetric threshold networks). In:
Proceedings of Cognitiva 85. Paris, France. 4–7 June, 1985. pp. 599–604.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

254 Bibliography

217. Rumelhart, D.E., Hinton, G.E., and Williams, R.J., 1986. Learning repre-
sentations by back-propagating errors. Nature, 323, pp. 533–36.
218. Hornik, K., Stinchcombe, M., and White, H., 1989. Multilayer feedforward
networks are universal approximators. Neural Networks, 2(5), pp. 359–66.
219. Ng., A., 2018. Heroes of Deep Learning: Andrew Ng interviews Yann
LeCun. YouTube [online]. https://www.youtube.com/watch?v=
Svb1c6AkRzE. (Accessed 14 August 2019).
220. LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E.,
Hubbard, W., and Jackel, L.D., 1989. Backpropagation applied to handwrit-
ten zip code recognition. Neural Computation, 1(4), pp. 541–51.
221. Thorpe, S., Fize, D., and Marlot, C., 1996. Speed of processing in the human
visual system. Nature, 381(6582), pp. 520–2.
222. Gray, J., 2017. U of T Professor Geoffrey Hinton hailed as guru of new
computing era. The Globe and Mail, April 7.
223. Allen, K., 2015. How a Toronto professor’s research revolutionized artifi-
cial intelligence. The Star. April 17.
224. Hinton, G.E., Osindero, S., and Teh, Y.W., 2006. A fast learning algorithm
for deep belief nets. Neural Computation, 18(7), pp. 1527–54.
225. Ciresan, D.C., Meier,U., Gambardella, L.M., and Schmidhuber, J., 2010.
Deep big simple neural nets excel on handwritten digit recognition. arXiv
preprint arXiv:1003.0358.
226. Jaitly, N., Nguyen, P., Senior, A.W., and Vanhoucke, V., 2012. Application of
pretrained deep neural networks to large vocabulary speech recognition.
In: Proceedings of the 13th Annual Conference of the International Speech Communication
Association (Interspeech). Portland, Oregon, 9–13 September 2012. pp. 257–81.
227. Hinton, G., et al., 2012. Deep neural networks for acoustic modeling in
speech recognition: The shared views of four research groups. IEEE Signal
Processing Magazine, 29(6), pp. 82–97.
228. Krizhevsky, A., Sutskever, Il, and Hinton, G.E., 2012. ImageNet classi-
fication with deep convolutional neural networks. In: C. Burges, ed.,
Proceedings of the 27th Annual Conference on Neural Information Processing Systems
2013. 5–10 December 2013, Lake Tahoe, NV. Red Hook, NY Curran.
pp. 1097–1105.
229. Bengio, Y., Ducharme, R., Vincent, P., and Jauvin, C., 2003. A neural prob-
abilistic language model. Journal of Machine Learning Research, 3, pp. 1137–55.
230. Sutskever, I., Vinyals, O., and Le, Q.V., 2014. Sequence to sequence learning
with neural networks. In: Z. Ghahramani, M. Welling, C. Cortes, N.D.
Lawrence, and K.Q. Weinberger, eds., Proceedings of the 28th Annual Conference
on Neural Information Processing Systems 2014. 8–13 December 2014 , Montreal,
Canada. Red Hook, NY Curran. pp. 3104–12.
231. Cho, K., Van Merriënboer, B., Bahdanau, and Bengio, Y., 2014. On the
properties of neural machine translation: Encoder-decoder approaches.
arXiv preprint arXiv:1409.1259.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 255

232. Bahdanau, D., Cho, K, and Bengio, Y., 2014. Neural machine trans-
lation by jointly learning to align and translate. arXiv preprint arXiv:
1409.0473.
233. Wu, Y, et al., 2016. Google’s neural machine translation system: Bridg-
ing the gap between human and machine translation. arXiv preprint
arXiv:1609.08144.
234. Lewis-Kraus, G., 2016. The great A.I. awakening. The New York Times,
December 20.
235. LeCun, Y, Bengio, Y., and Hinton, G., 2015. Deep learning. Nature.
521(7553), pp. 436–44.
236. Vincent, J., 2019. Turing Award 2018: Nobel prize of computing given to
‘godfathers of AI’. The Verge [online]. https://www.theverge.com/2019/3/
27/18280665/ai-godfathers-turing-award-2018-yoshua-bengio-geoffrey-
hinton-yann-lecun. (Accessed 19 May 2020).
237. Foster, R.W., 2009. The classic of Go. http://idp.bl.uk/. (Accessed 20 May
2020).
238. Moyer, C., 2016. How Google’s AlphaGo beat a Go world champion.
The Atlantic, March.
239. Morris, D.Z., 2016. Google’s Go computer beats top-ranked human. For-
tune, March 12.
240. AlphaGo, 2017 [film]. Directed by Greg Kohs. USA: Reel as Dirt.
241. Wood, G., 2016. In two moves, AlphaGo and Lee Sedol redefined the
future. Wired, March 16.
242. Metz, C., 2016. The sadness and beauty of watching Google’s AI play Go.
Wired, March 11.
243. Edwards, J., 2016. See the exact moment the world champion of Go
realises DeepMind is vastly superior. Business Insider [online]. https://
www.businessinsider.com/video-lee-se-dol-reaction-to-move-37-and-
w102-vs-alphago-2016-3?r=US&IR=T. (Accessed 19 May 2020).
244. Silver, D., et al., 2016. Mastering the game of Go with deep neural net-
works and tree search. Nature, 529(7587), pp. 484–9.
245. Hern, A., 2016. AlphaGo: Its creator on the computer that learns by
thinking. The Guardian, March 15.
246. Burton-Hill, C., 2016. The superhero of artificial intelligence: can this
genius keep it in check? The Guardian, February 16.
247. Fahey, R., 2005. Elixir Studios to close following cancellation of key
project. gamesindustry.biz [online]. https://www.gamesindustry.biz/
articles/elixir-studios-to-close-following-cancellation-of-key-project.
(Accessed 19 May 2020).
248. Mnih, V., et al., 2015. Human-level control through deep reinforcement
learning. Nature, 518(7540), p. 529.
249. Silver, D., et al., 2017. Mastering the game of Go without human knowl-
edge. Nature, 550(7676), pp. 354–9.
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

256 Bibliography

250. Silver, D., et al., 2018. A general reinforcement learning algorithm

that masters Chess, Shogi, and Go through self-play. Science, 362(6419),
pp. 1140–4.
251. Lovelace, B., 2018. Buffett says cryptocurrencies will almost certainly
end badly. CNBC [online]. https://www.cnbc.com/2018/01/10/buffett-
says-cyrptocurrencies-will-almost-certainly-end-badly.html. (Accessed
19 May 2020).
252. Anonymous, n.d. Market capitalization. blockchain.com [online]. https://
www.blockchain.com/charts/market-cap. (Accessed 19 May 2020).
253. Hughes, E., 1993. A Cypherpunk’s manifesto – Activism. https://www.
activism.net/cypherpunk/manifesto.html. (Accessed 19 May 2020).
254. Assange, J., Appelbaum, J., Maguhn, A.M., and Zimmermann, J., 2016.
Cypherpunks: Freedom and the Future of the Internet. London: OR books.
255. Driscoll, S., 2013. How Bitcoin works under the hood. Imponder-
able Things [online]. http://www.imponderablethings.com/2013/07/how-
bitcoin-works-under-hood.html#more. (Accessed 19 May 2020).
256. Nakamoto, S., 2008. Bitcoin: A peer-to-peer electronic cash system. Work-
ing Paper.
257. Webster, I., 2020. Bitcoin historical prices. in2013dollars.com [online].
http://www.in2013dollars.com/bitcoin-price. (Accessed 22 June 2020).
258. Anonymous, n.d. Bitcoin all time high – how much was 1 bitcoin worth
at its peak? 99BitCoins [online]. https://99bitcoins.com/bitcoin/historical-
price/all-time-high/#charts. (Accessed 19 May 2020).
259. Anonymous, 2019. Bitcoin energy consumption index. https://
digiconomist.net/bitcoin-energy-consumption. (Accessed 19 May 2020).
260. L.S., 2015. Who is Satoshi Nakamoto? The Economist, November 2.
261. Greenberg, A., 2016. How to prove you’re Bitcoin creator Satoshi
Nakamoto. Wired, April 11.
262. Feynman R.P., 1982. Simulating physics with computers. International Journal
of Theoretical Physics, 21(6), pp. 467–88.
263. Shor, P.W., 1982. Polynomial-time algorithms for prime factorization
and discrete logarithms on a quantum computer. SIAM Review, 41(2),
pp. 303–32.
264. Anonymous, n.d. Quantum – Google AI. https://ai.google/research/
teams/applied-science/quantum-ai/. (Accessed 19 May 2020).
265. Dyakonov, M., 2018. The case against quantum computing. IEEE Spectrum,
November 15.
266. Arute, F., et al., 2019. Quantum supremacy using a programmable super-
conducting processor. Nature, 574(7779):505–10.
267. Savage, N., 2019. Hands-on with Google’s quantum computer. Scientific
American, October 24. (Accessed 19 May 2020).
 OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi

Bibliography 257

268. Anonymous, Machine translation and automated analysis of cuneiform

languages (MTAAC Project). https://cdli-gh.github.io/mtaac/. (Accessed
22 June 2020).
269. Nemiroff, R. and Bonnell J., 1994. The square root of two to 10 million dig-
its. https://apod.nasa.gov/htmltest/gifcity/sqrt2.10mil. (Accessed 19 May
2020).
270. Keough, B., 1997. Guide to the John Clifford Shaw Papers. http://sova.si.
edu/record/NMAH.AC.0580. (Accessed 19 May 2020).
271. Christofides, N., 1976. Worst-case analysis of a new heuristic for the trav-
elling salesman problem. Technical report, DTIC Document.
272. Sebo, A. and Vygen, J. 2012. Shorter tours by nicer ears. arXiv preprint
arXiv:1201.1870.
273. Mullin, F.J. and Stalnaker, J.M., 1951. The matching plan for internship
appointment. Academic Medicine, 26(5), pp. 341–5.
274. Anonymous, 2016. Lloyd Shapley, a Nobel laureate in economics, has died.
The Economist, March 13.
275. Merkle, R.C., 1978. Secure communications over insecure channels. Com-
munications of the ACM, 21(4), pp. 294–99.
OUP CORRECTED PROOF – FINAL, 15/7/2020, SPi
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index

π ensemble weather forecast 71

definition 25–26 etymology 31
enumeration 26–30 Euclid’s, iterative 20
mnemonic 25 Euclid’s, recursive 21
Fourier transform 36
A* algorithm 106–107 genetic 113
abacus 39 Halting Problem 50–51
Abbe, Cleveland 56 Insertion Sort 4
Aberdeen Proving Ground 59, 63, 68 minimax 89
Adelard of Bath 31 Monte Carlo 66
Adleman, Leonard 136–137, 141 Netflix recommender 165
photograph 136 numerical weather forecast 58
Adobe 200 packet routing 123
Afghanistan 18 PageRank 229
AI Winter 92, 188 position evaluation 88
Aiken, Howard 54 proof by contradiction 50
Akkad public key encryption 136
Empire 10–11 Quicksort 5
language 10–12 RSA decryption 138
al-Aas, General Amir ibn 23 RSA encryption 138
al-Khwārizmī, Muhammad ibn RSA key generation 139
Mūsā 30–31 Shor’s Algorithm 232
al-Rashid, Caliph Harun 30 Sieve of Eratosthenes 22
Alexander the Great 18 square root of two 16
Alexandria 18–20, 22 Travelling Salesman Problem 93–94
Library of 19, 21–23 Turing machine 48
Lighthouse of 19 algorithms
map 9 growth of 73
Museum of 19, 23 importance of 7–8
Temple of Serapis 23 Algoritmi de Numero Indorum 31
AlexNet 198 Alphabet 158
algebra 30, 82, 84–85 AlphaGo 203–210, 212
algorithm artificial neural network 207–208
Amazon recommender 150 hardware 207
Archimedes’ 28 performance 209–210
artificial neural network 186 training 208–209
artificial neural network training 230 AlphaGo Zero 212–213
Bitcoin 231 AlphaZero 213
cistern 13 Amazon 147–149, 151, 158, 162, 200
decision-making 2 Alexa 36
definition 1 Analytic Engine 42–45
divide by two 3 Andreessen, Marc 146
divide even number by two 2 Apollo 11 119
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

260 Index

Apple 124, 200 Automatic Computing Engine

I 143 (ACE) 75–76, 120
II 144 Pilot 76, 77
Macintosh 144
Siri 36 Babbage, Charles 40–45, 54
Arabic 30, 31 portrait 41
Archimedes 26–30 Babylon 11
Archimedes’ algorithm 27–29 Babylonia 14, 17, 26
Aristotle 55 early mathematics 15
ARPA (Advanced Research Projects Empire 11
Agency) 117, 118, 124 language 12
IPTO (Information Processing Old period 13
Techniques Office) 118–119, 124 weather forecasts 55
ARPANET 119–127 Backgammon 211
first message 119–120 BackRub 156
artificial intelligence (AI) 79, 91–92, 228 Baidu 200
board games 79 Ballistic Research Laboratory (BRL) 59, 63
can machines think? 81–82 Baran, Paul 120
definition 80 Battleships 102
human-level general (HLAGI) 80–81, Bechtolsheim, Andy 156–157
200–201 Behistun Inscription 12
artificial neural network (ANN) 178, Bell Labs 86, 129–130, 133, 192, 194, 196, 199
183–201, 207–213 Bell, Alexander Graham 129
ability to generalize 191 BellKor’s Pragmatic Chaos 167–168
activation function 184, 189–190 Bells Labs 194
autoencoder 200 Bengio, Yoshua 198–200
back-propagation 189–191 photograph 198
bias 184 quote 195
drop-out 198 Berners-Lee, Tim 144–146, 158
excitation 184 Bethe, Hans
forward-propagation 190 quote 69
fully connected 185 Bezos, Jeff 146–147, 158
hidden layer 185 quote 147
inference 190 Bezos, MacKenzie 147
unsupervised pretraining 196 big data 163, 168, 169
weight 184 binary numbers 130
artificial neuron Bitcoin 215, 217, 224
diagram 185 mining 221–222
Ashurbanipal, King 17 transaction diagram 218
Association of Computing Machines Bjerknes, Vilhelm 56
(ACM) 53 Black Sea 11
Assyria 11 Bletchley Park 51–53, 76, 141
Astronomical Society 41 Blias, Frances 60
AT&T 194, 199 Block, HD
Atanasoff, John 63 quote 188
Atanasoff–Berry Computer (ABC) 63 Blockchain 217–222
Atari 211 diagram 220
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index 261

Bolt, Beranek, and Newman Technologies Charney, Jule 68–69

(BBN) 119, 124, 127 Checkers 76–81, 91
Bombe 52 board 78
Bonaparte, Napoléan 33 rules 77–78
Bosporus 11 Samuel 86–91
Boston Pool algorithm 109–110 Strachey 78
Brazil 70 Chess 91, 204, 210, 211, 213
Brin, Sergey 151–152, 156–158 China 207
photograph 152 abacus 39
quote 151–152 early mathematics 29
British East India Company 12 Yangtze 29
British Museum 18 Yellow River 29
British Navy 52 Church, Alonzo 51
Brown University 107 Cinematch 163, 164, 167, 168
bug 44 ciphertext 134
Burges, Christopher 196 circle 25
Burks, Arthur 62–63, 111 polygon approximation 27–29
Bush, Vannevar 145 diagram 27
quote 143 circuit-switching 120
Butterfly Effect 70, 71 circumference 25
Byron, Lady 42–43 of the Earth 22
Byron, Lord 42–43 cistern problem 13–14
Clark, Wesley 183
Caesar Cipher 133–134 Clarke, Joan 52
Caesar, Julius 23, 133 class 184
Cailliau, Robert 145 classifier 184
Cajal, Santiago Ramón y 180 Clay Mathematics Institute 102
calculator Clement, Jack 42
mechanical 39–40 clock arithmetic 138–139
Caltech 224 Cocks, Clifford 141
Canada 180, 192 Cold War 37, 91, 115, 117
Montreal 199 combinatorial optimization
Toronto 101, 199 quantum 225–226
Canadian Institute For Advanced Research combinatorial optimization 102–103, 107,
(CIFAR) 195 110
Canfield Solitaire 64–65 Compendious Book on Calculation by Completion and
diagram 65 Balancing 30
Carnegie Institute of Technology 83–85 complexity classes 99–102
Carnegie Mellon University 85, 91, >NP 100–102
189, 195 diagram 99
case-based reasoning 177–178 NP 99–100
Centers for Disease Control and Prevention NP-Complete 101–102
(CDC) 168–170 NP-Hard 101–102
Cerf, Vint 125–127 NP\P 100–102
CERN (European Organization for Nuclear P 99–100
Research) 144–146 P versus NP 100–102
Chandler, Raymond 148 table 99
Chaos Theory 70–71 computability 49–51
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

262 Index

computational complexity 97–102 Dartmouth Conference 80, 82

constant 97–98 Davies, Donald 120
definition 6 Dawkins, Richard 113–114
graph 97 DE Shaw & Company 147
linear 98 decision-making 2–3, 17–18
polynomial time 98 deduction 84
quadratic 98 Deep Blue 170
quasilinear 98 deep learning 194–201
superpolynomial time 98 DeepMind Technologies 204–206, 210–213
computer 228 diameter 25
Chess 91 Difference Engine 41–42, 54
chip 73 Diffie, Whitfield 135–136, 140–141
electronic 224–225 digit recognition 193–194, 196–197
first music 77 Digital Equipment Corporation (DEC) 124,
first operational electronic 59–64 144
general-purpose 7, 49 digital signature 218–219
Go 203 diagram 218
mainframe 75 Dijkstra’s algorithm 104–106
microcomputer 143 Dijkstra, Edsger 103–107, 110, 126
minicomputer 143 photograph 103
operation of 6–7 distributed algorithm 122
quantum 224–228 divide-and-conquer 6
symbol manipulator 75 division 14–15
computer science 115 division by two 1–3
condition 2 DNA 113
convolutional neural network 193–194, domain name system 126
197–198 dotcom bubble 152, 157–158
diagram 193 Draughts 76
Cook, Stephen 100–102
Cooley, James 37 eBook 118
Cornell University 183, 224 Eckert, Presper 59–63
Corporation for National Research École Polytechnique 33
Initiatives (CNRI) 127 École Supérieure d’Ingénieurs en
correlation 35 Electrotechnique et Electronique
Cortes, Corinna 196 (ESIEE) 192
cryptocurrency 215–224 Edinburgh University 91, 195
Double-Spend Problem 216–217 Edison, Thomas 45
cuneiform 10–12 EDVAC 60–62
cybernetics 182 Egypt 18, 26, 33, 133
Cypherpunks 215–216, 223 Alexandria 18–20, 33
A Cypherpunk’s Manifesto 216 Aswan 22
Cyrene 21 Cairo 33
CYRUS 177 early mathematics 18
Cyrus the Great 11 Giza 31
hieroglyphics 18
DAB (Digital Audio Broadcasting) 37 invasion 33
Darknet 221 Memphis 18
Dartmouth College 80 Einstein, Albert 61
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index 263

Eisenhower, President Dwight D. 117 Facemash 159

Elamite language 12 Farley, Belmont 183
Elements of Geometry 19–20 fast Fourier transform 37
Elixir Studios 211 Fee, Greg 17
Ellis, James 141 Feigenbaum, Edward 177
quote 141 Fellow of the Royal Society 41
email 118 Ferranti Mark I 79, 145
encryption 133 Ferrucci, Dave 172, 174
asymmetric keys 135–136 Feynman, Richard 224
attack 134 Fibonacci 31
key 134 Fields Medal 101
Key Distribution Problem 134–135 First Draft of a Report on the EDVAC 62
symmetric keys 135 First World War 168
Engelbart, Douglas 144 Fisher, Ronald 111
ENIAC 59–64, 68, 72, 82, 111 Forbes 158, 162
demonstration 62–63 Fourier Series 33–34
numerical weather forecasts 68–69, 73 Fourier transform 34–37, 37, 197, 226
patent 63 quantum 227
photograph 60 Fourier, Jean-Baptiste Joseph 32–37
programmers 60, 63 bust 32
Enigma 52, 53 fractional point 14–15
Enlightenment 32 France 39
ensemble method Auxerre 32
recommender systems 167 Eiffel Tower 37
weather forecast 71–72 Grenoble 33
map 71 Paris 33, 189, 191, 192, 199
Epic of Gilgamesh 9, 12, 55 French Revolution 32, 33
Epstein, Edward 71–72
equilateral 27 Göbekli Tepe 25
Eratosthenes 21–23 Gale, David 107–110
error correction 127–133, 227 Gale–Shapley algorithm 108–110
checksum 128–129 game tree
parity 130–133 Checkers 78–79
redundancy 127–128 Go 205
Ethereum 222 Gardner, Martin
Ethernet 124 quote 140
Euclid 19–21 Gauss, Carl Friedrich 37
Euclid’s algorithm 20–21, 226, 227 General Problem Solver 85
Euphrates 9, 25 genetic algorithm 111–115
Eureka 26–27 The Genetical Theory of Natural Selection
European Centre for Medium-Range 111
Weather Forecasts, 72 Germany 37, 39
expert systems 176–178 Berlin 12, 53
Go 91, 203–213
Facebook 159–163, 168, 200, 222 complexity 205
EdgeRank 161–162 rules 204
Libra 222 Goldstine, Adele 60
News Feed 161–163 Goldstine, Herman 59–62
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

264 Index

Google 126, 127, 156–158, 168, 169, 197–198, Hindu–Arabic numerals 30–31
200, 204, 206, 227–228 On the Hindu Art of Reckoning 30–31
AdWords 157 Hinton, Geoffrey 192, 194–200
Cloud 30 photograph 195
DeepMind Challenge 203 Hoare, Tony 4
Flu Trends 168 Holberton, Betty 60
Labs 196 Holland 103
PageRank 152–157 Amsterdam 103
Translate 199–200 Holland, John 110–115
Government Codes and Ciphers photograph 110
School 52 quote 111
Government Communications Holmes, Sherlock 148
Headquarters (GCHQ) 140–141 Horn, Paul 172
Graph Colouring Problem 102 Hot Or Not 159
graphical user interface (GUI) 124, 144 House of Wisdom 30–31
greatest common divisor (GCD) 20–21 Huang, Aja 203, 205, 206
Greece 133 Hughes, Eric
Athens 22 quote 216
early mathematics 15, 18, 23 Hui, Fan 206, 212
language 30 quote 206
greyscale 179 Hui, Liu 29
human brain 180–181
Halting Problem 49–51 Hungary
diagram 50 Budapest 61
Hamilton, William 93 Hypatia 23
Hamming, Richard 129–130, 132–133
photograph 129 IBM 54, 73, 86, 110, 124, 125, 129, 135, 144,
quote 130 170, 178, 183, 200, 227
Hammurabi, King 11, 13 Jeopardy Challenge 171–178
Code of 17 PC 144, 146
hardware 7 Watson 171–178
Harmon, Leon 61 quote 173–174
harmonics 33–34, 36–37 Watson Center 135, 171
Harrow School 76 image recognition 197–198
Harvard Mark I 54 ImageNet 198
Harvard University 54, 64, 110, 118, 159, 189 India 31, 45, 133
hashing algorithm 219–220 language 30
Hassabis, Demis 205, 210–212 Industrial Revolution 37,
photograph 210 39–40, 112
quote 211 infinite loop 49
Hastings, Reed 163, 168 influenza 168–170
Hebb, Donald 180 Insertion Sort 3–4, 6
Hellman, Martin 134–136, 141 computational complexity 98
Helsgaun, Keld 97 diagram 3
Herbert, Simon 82–86 Institute for Advanced Study (IAS)
photograph 83 61, 64, 68
Heron of Alexandria 16 computer 67–68, 83
Heron’s algorithm 16 integrated circuit 72–73, 228
Hidden Markov Model (HMM) 197–198 Intergalactic Computer Network 118
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index 265

Intel 73 Korea 204

International Computer Communication Krakauer, David
Conference (ICCC) 124, 125 quote 115
International Standard Book Number Krizhevsky, Alex 198
(ISBN) 129
Internet 118, 124, 127, 133, 141, 144, 147, 158 Lambert, Johann Heinrich 26
first message 126 Larson, Judge 63
Society 127 Latin 31
Iowa State University 63 LeCun, Yann 189, 191–194, 197,
Iran 12 200–201
Iraq photograph 192
Baghdad 12, 30, 31 quote 200–201
Ireland 73 Lee, Sedol 203–207, 212
irrational number 26 quote 206
Islamic Golden Age 30 Legg, Shane 211
Israel Lettvin, Jerome 182, 183
Tel Aviv 137 quote 182
Italy Liar’s Paradox 50
Mount Vesuvius 45 Liber Abaci 31
Torino 42 Libraries of the Future 118
iteration 1–2 Lichterman, Ruth 60, 61
Iwao, Emma Haruka 30 Licklider, JCR 118–119, 124, 127
photograph 118
Jacquard’s Loom 39–40, 42 quote 117
Jacquard, Joseph Marie Charles 39 Lighthill, James 92
Jaitly, Navdeep 197–198 Linden, Greg 147–149, 158, 162
Japan 30, 204 photograph 148
Jennings, Jean 60, 61 Lionhead Studios 210
Jennings, Ken 171–174 Logic Theorist 84
quote 171–174 lookahead 78–79, 89–90
Jeopardy 171 loop 6
Jie, Ke 212–213 Lorenz, Edward 69–70, 72
Jobs, Steve 143 love letter 77
JOHNNIAC 83, 84 Lovelace, Ada 42–45
Jones, William 26 portrait 41
Lucasian Professor of Mathematics 41, 92
Kahn, Bob 124–127 Luxor 18
quote 125 Lviv Polytechnic Institute 64
Kamp, Richard 102 Lyceum 55
quote 102 Lynch, Owen 73
Kasparov, Garry 170 Lynch, Peter 73
King, William 43
Kirkman, Thomas 93 machine learning 86–89, 91, 163, 168–170,
Kleiner Perkins 157 188
Kleinrock, Leonard 119 definition 86
Kline, Charley 119–120 recommender 166–167
Knapsack Packing Problem 102 Manchester Baby 76–77
Kolodner, Janet 177 Manchester Mark I 79
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

266 Index

Manchester University Computer Monte Carlo method 64–67, 71

52, 77 Monte Carlo tree search 207
quote 77 Moore School 59–60, 62, 63
Man-Computer Symbiosis 118 Moore’s Law 73
Manhattan Project 62, 64, 129, 224 Moore, Gordon 73
Mansfield Amendments 92 Morcom, Christopher 46
Mark I NPL Network 120 Mosaic 146
Massachusetts Institute of Technology move 78
(MIT) 70, 86, 91, 110, 118, 119, 124, movie ratings
135, 136, 141, 182, 183, 187, 189, table 165
199, 224 Murray, Arnold 52
Lincoln Labs 83 Muskie, Edmund 177
matching problem 107 MYCIN 177
Mathematisch Centrum 103
Mauchly, John 59–63 Nakamoto, Satoshi 217–223
quote 82 quote 222
McCarthy, John 80–82 Napoleon Dynamite problem 167
McCulloch, Warren 182–183 NASA (National Aeronautics and Space
McGill University 199 Administration) 114, 115, 117,
McKinsey 151, 168 119, 127
McNulty, Kathleen 60 NASDAQ 157
Mediterranean 19, 23, 30 National Physical Laboratory (NPL) 52,
Mellon Institute 85 75–76, 120
memex 143 National Research and Development
memory 17 Corporation (NRDC) 77
Menabrea and Lovelace National Residency Matching Program
quote 43–44, 75 (NRMP) 109–110
Menabrea, Luigi Federico 42–45 National Security Agency (NSA) 135, 140
Merkle, Ralph 135, 141 National University of Singapore 196
Mesopotamia 9–18, 31, 133, 228 natural evolution 112
map 9 natural language processing 197
Meteorologica 55 Bengio 198–199
Meteorological Office (UK) 55 IBM Watson 175
Metropolis, Nicholas 64, 67 translation 199–200
quote 64 Nature 168, 189, 200, 211
Microsoft 124, 147, 158, 222, 227 Naval Postgraduate School 133
Research 196 NCP 126
Windows 124, 144, 146 Nealey, Robert 90
Middle East 11, 18 Netflix 163–168
Millennium Problems 102 Prize 163–166, 168–170
minimax 89–90 Netscape 152, 157
algorithm 207 neutron penetration problem 64, 66–67
diagram 90 New York Times 184
Minsky and Papert New York University 134, 194
quote 188 Newcastle University 56
Minsky, Marvin 80, 91, 187–189 Newell, Allen 82–86, 91
MITRE Corporation 135 photograph 83
MNIST dataset 196–197 quote 83–84
model 58 Newman, Max 77
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index 267

Ng, Andrew 200 Perceptrons 187

Nile 19 period 33
Nine Chapters on the Mathematical Art 29 Persia 18, 31
Nineveh 17 Empire 11, 18
Nobel Prize 86, 101, 110, 180, 224 language 12
Nokia 6300 73 Persian Gulf 9, 11
non-secret encryption 141 personalization 148, 161,
Norway 126 162, 163
NSFNET 127 PGP 216
nuclear bomb 62 PhD 51
tests 37 Phoenicia 18
numerical weather forecasts 56–59, PHONIAC 73
67–72 Pitts, Walter 181–183
pixel 179
Official Secrets Act 141 plaintext 134
omens 17 ply 78
Oppenheimer, Robert 61 Poker 211
optimization 112, 115 Poland 52
Osindero, Simon 196 Warsaw 64
position evaluation 86–89
packet 120–122 prediction error 164
routing 122–123 premises 84
table 123 prime factorization 223–224, 226
packet-switching 120–124 prime numbers 22–23
diagram 120 enumeration 23
Page, Larry 151–152, 156–158 Princeton University 51, 61, 62, 64, 85, 107,
photograph 152 124, 147, 187
quote 151 Principia Mathematica 85, 181
PageRank 152–157 program 7
link table 153 Project MAC 118
mini-WWW 153 proof 84
diagram 153 protocol 125–126, 146
table 155 Ptolemy
Pakistan 11 I 18–19
Indus Valley 31 III 21–22
Papert, Seymour 187–189, 192 XIV 23
papyrus 18 public key cryptography 135–141, 217–218
paradox 50 Pythagorean Theorem 15, 29
parity bit 130
Parker, David 189 quantum computer 224–228
pattern recognition 179–180, 188, interference 225
197, 210 noise 227
PC Magazine 157 qubit 225–227
Pentagon 119 state collapse 225
Pentamind 211 supremacy 227–228
pepper moth 112 Queen of England 53–158
Perceptron 183–188 Quicksort 4–6
diagram 186 computational complexity 98
multilayer 189 diagram 5
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

268 Index

radar 76, 110 Santa Fe Institute 115

radius 25 satoshi 239
RAND Corporation 82–84, 108 Schmidhuber, Jürgen 197
Randolph, Marc 163, 168 Science Research Council 92
Rawlinson, Sir Henry 12 Scientific American 140
recommender system scientific paper 41
Amazon 148–151 Scrabble 211
factor 164–165 search
product similarity table 149–150 exhaustive 85, 93–97
training 166–167 heuristic 85, 90
recursion 21 Sebokht, Bishop Severus 31
Red Sea 19 Second World War 59, 62, 64, 82, 86, 129, 140
reinforcement learning 88, 91, 209 Secure Socket Layer 140
Remington Rand 63 Sedol, Lee
Renaissance 31–32 God’s move 206
Rensselaer Polytechnic Institute 172 quote 206
Rhind Papyrus 18, 26, 29 Selfridge, Oliver 83
Rhind, Henry 18 sequential 1
Richardson, Lewis Fry 56–59, 69, 71 Sequoia Capital 157
photograph 57 sexagesimal 14
The Richest 158 Shakey the robot 106–107
Rivest, Ronald 136–137, 141 Shamir, Adi 136–137, 141
photograph 136 photograph 136
Roberts, Larry 119 Shannon, Claude 80, 91
Rockefeller Foundation 80, 82 Shapley, Lloyd 107–110
Roman numerals 31 Shogi 213
Rome Shor’s algorithm 224, 226–227
abacus 39 Shor, Peter 224
ancient 29, 30 Sicily
Empire 30 Syracuse 26, 29
Rosenblatt, Frank 183–188 Sieve of Eratosthenes 22–23
photograph 183 signal 36–37
Roth, Alvin 110 Silicon Valley 157
routing-finding 103–107 Silver, Don 205
Royal College of Surgeons 235 Simon, Herbert 82, 91
RSA 136–140, 223 Simulating Physics with Computers 224
decryption 137–139 simulation 58
encryption 137–139 Sinclair Spectrum 48K 210
key generation 139–140 Sketch of the Analytical Engine 43
Russell, Bertrand 181 SNARC 188
Russia social network 159–160
Moscow 151 software 7
Rutter, Brad 171–174 Some Studies in Machine Learning using the Game of
quote 173–174 Checkers 86
sorting 3–6
sampling 35 South Africa 187
Samuel, Arthur 86–91, 207 South Korea 204
photograph 86 Seoul 203, 204
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index 269

Spain 180 Times 223

speech recognition 197–198 topology 187
Spence, Homer 61 Tor 216
Spotify 36 trace 44
Sputnik 1, 117 transistor 72–73
square root of two 16–17 Travelling Salesman Decision
enumeration 17 Problem 101–102
Sri Lanka 4 Travelling Salesman Problem 97, 101–102
Stable Marriage Problem 107–110 computational complexity 96–97
table 109 definition 93
Stabo diagram 93
quote 19 exhaustive search 93–97
Stanford Research Institute (SRI) 106, quantum 225–226
119–144 tree diagram 95
Stanford University 83, 91, 110, 125, 135, world record 97
141, 151, 156, 157, 177 Trebek, Alex 173–174
state 47 quote 173–174
Station X 52, 53 Tukey, John 37
Stephenson, Robert 42 Turing Award 53, 85, 107, 127, 141, 158,
Strachey, Christopher 76–79, 79, 86 188, 200
photograph 76 Turing Complete 49–54
Suleyman, Mustafa 211 Turing machine 47–49, 53, 182
Sumer 9–11, 25 diagram 47
abacus 39 Turing Test 52, 53
Eridu 9 Turing, Alan 45–53, 75–77, 81–82, 120
Kish 9 Halting Problem 49–51
language 10–12 photograph 46
Ur 9 quote 81
Uruk 9 Turing, Ethel Sara 46
supervised training 196 quote 46
Sutskever, Ilya 198 Turkey 9, 25
Switzerland 197 Twitter 200
Zurich 61
Syene 22 U-Boat 52
symbolic logic 188 UK 126, 223
symbolic reasoning 85 Cheltenham 141
Syria 18 Eskdalemuir 56–57
Hucknall 45
Taiwan 203 London 12, 43, 45, 52, 76, 120, 210
Taylor, Bob 118–119, 124–144 Newcastle 56
quote 119 Walworth 40
TCP/IP 126–127 Wimbledon 194
Teh, Yee-Whye 196 Ukraine 64
Teller, Edward 61 Ulam, Stanislaw 64–67, 69, 71
Theophilus of Alexandria, Pope 23 photograph 64
Theophrastus 55 uniform resource locator (URL) 145,
Théorie Analytique de la Chaleur 33 152, 156
Tidal 36 Université de Montréal 195, 199
Tigris 9 Université Pierre et Marie Curie 192
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

270 Index

University College Dublin 73 Fort Wayne 110

University College London 195, 211 Los Alamos 64, 69, 129, 224
University of Budapest 61 Milwaukee 83
University of California, Berkeley 101, 102, Montrery 133
119, 135 New Haven 125
University of California, Los Angeles New Jersey 192
(UCLA) 119, 125 New Rochelle 183
University of California, San Diego 189, 195 New York 71, 134, 137, 187
University of Cambridge 41, 46, 51, 52, 75, Pennsylvania 54
92, 181, 187, 194, 211 Pittsburgh 83, 84
King’s College 56, 76, 77 San Francisco 83, 126, 137
University of Chicago 181 Santa Fe 69
University of Geneva 187 Santa Monica 84
University of Illinois 86, 129, 182 Seattle 147
University of Illinois at St. Louis 118
Urbana-Champaign 146 Texas 70
University of Leiden 103 White Plains 159
University of Manchester 52, 76 Yorktown 171
Computing Machine Laboratory 77 USSR 37, 117
University of Maryland 151
University of Michigan 60, 111, 115, 151 van Wijngaarden, Adriaan 103, 107
University of Oxford 4, 79, 144 Vance, Cyrus 177
University of Paris 187 Vietnam War 92
University of Pennsylvania 59 viral messaging 162
University of Rochester 118 VisiCalc 143–144
University of Southern California 126 von Neumann, John 59, 61–63,
University of Stockholm 71 67–182
University of Sussex 195 photograph 68
University of Toronto 101, 192, 194–196
University of Washington 148 Wales 26
unsupervised training 196 Wall Street 146, 147, 152
Ursinus College 59 Walmart 158
US Air Force 71, 82, 110 waveform 33–36
US Army 54, 59, 69 amplitude 34
US Department of Defense 115 compound 34
US Navy 59, 67, 110, 184 superpose 34
US 37, 64, 117, 126, 151, 158, 160, 168, 169, Weather Bureau (US) 55, 56, 68
171, 181, 194, 216 Weather Prediction by Numerical Process 58
Albuquerque 147 On Weather Signs 55
Berkeley 135 web
Boston 64, 183 browser 145
Cambridge 102, 119 crawler 156
Chesapeake Bay 188 hyperlink 145
Chicago 129 page 145
Cincinnati 59 search 152–156
Connecticut 69 server 145
Dallas 118 Welchman, Gordon 52
Dartmouth 80 Werbos, Paul 189
Detroit 181 quote 189
 OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi

Index 271

Wescoff, Marlyn 60 Yahoo! 200

Whirlwind 110 Yale University 15, 177
Wiener, Norbert 182, 183 YBC 7289 16
WiFi 37 photograph 15
Williamson, Malcolm 141
World Wide Web 140, 144–146, 151, Zagros 12
156, 158 zero 14–15, 31
World Wide Web Consortium Zhoubi Suanjing 29
146 Zu Chengzhi 29
Wozniak, Steve 143 Zuckerberg, Mark 159–162
photograph 160
Xerox PARC 124–144 Zuse, Konrad 53–54
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi
OUP CORRECTED PROOF – FINAL, 13/7/2020, SPi