Reasoning

Classical AI
• Humans like to compare ourselves with others

• We try to find ways in which we are better than someone or

something else

• As artificial intelligence was born, the desire arose to compare

artificial intelligence with human intelligence

• Basic rule was that human intelligence is as good as it gets -

even human intelligence is the only intelligence

• The best artificial intelligence can achieve is to be as good as

human intelligence - to copy it in some way.
 Minsky Definition

• Classical artificial intelligence techniques focus on

getting a machine to copy human intelligence.

• Marvin Minsky definition, “Artificial intelligence is the

science of making machines do things that would
require intelligence if done by men”.

• Definition side steps the concept of what intelligence

is - merely points to machines copying humans.
 Herb Simon

• Philosophy in 1957 described by Herb Simon

• “There are now in the world machines that

think, that learn and that create. Moreover, their
ability to do these things is going to increase
rapidly until … the range of problems they can
handle will be coextensive with the range to
which the human mind has been applied”.
 Top Down Approach
• An approach to artificial intelligence arose
along the lines of a psychiatrist

• Attempt to understand the human brain’s

processing from the outside

• Then attempt to build a machine to copy

that way of functioning
 Reasoning

• One key aspect of human intelligence is the

ability of the human brain to reason

• Given a number of facts, the human brain

makes a reasoned assumption about a
situation and decides on a conclusion

• E.g. if it is 7am and my clock alarm is

ringing then it is time to get up
 Expert Systems

• An expert system reasons about facts in a

specific domain and works like an expert’s brain

• Needs knowledge about the domain, rules (from

experts) to follow when new information occurs
and a way of communicating with a user

• Called rule based systems, knowledge based

systems or expert systems.
 MYCIN

• Early successful working system - a medical

system
• MYCIN contained 450 rules - claimed to be
better than junior doctors and as good some
experts.
• Built by interviewing large numbers of
experts who reported from experience.
• The rules reflected the uncertainties with
medical conditions.
 Rules

• Rules of the form:

• IF (condition) THEN (conclusion).
• Several conditions may need to exist for
the rule to fire. A rule may be:
• IF (condition1 and condition2 or
condition3) THEN (conclusion).
• Example: IF (sneezing and coughing or
headache) THEN (flu).
 Conflict Resolution
• There might be several possible conclusions that can
be drawn from the facts

• The system needs further rules for such instances –

this is conflict resolution

• In many situations several conditions are met but

only one conclusion is required

• Decision is needed - which rules takes precedence

 Conflict Resolution - Techniques

• Highest priority rule

• Highest priority conditions
• Most recent condition
• Most specific (most conditions)
• Context limiting – rules in groups – a rule
must belong to an active group
 Multiple Rules
• Expert systems involve rules which depend on
each other. Example: engine management:
• Layer 1 Rules:
• IF (start button pressed) THEN (start engine)
• IF (gear selection) THEN (engage gears)
• Layer 2 Rule:
• IF (engine started and gears engaged) THEN
(vehicle drive)
 Forward Chaining

• Set of facts apparent at a particular time

• These fire a number of rules
• Realizing facts which fire other rules
• So on until a goal is reached
• Working from input data to end goal is
forward chaining
• Discover what can be deduced from facts.
 Backward Chaining
• Expert systems can be used in reverse fashion
• When a goal has been achieved - rules are searched
to see what facts occurred for the system to deduce
its conclusion
• What facts we must input to the system to realize a
specific goal?
• E.g. what happened to cause the vehicle to drive?
• From backward chaining - the start button was
pressed and the gear selection had been made.
• Backward chaining good for system verification,
where the expert system must be safety critical and
cannot arrive at a ‘wrong’ conclusion.
 Expert System Advantages
• Easy to program (IF-THEN structure).
• Each rule is a separate entity with its own data to
fire and its own individual conclusion drawn
• Ideal for dealing with natural real world information
• System structure is separate from the data - same
expert system structure could be employed in very
different domains
• Can deal with uncertainty, e.g. 75% certain about
the conclusion
• Speed of response – compared to a human expert
 Problems
• Gathering rules can be awkward. It is difficult for a person to
put into simple terms what it is they do in an everyday
situation. If several experts are being asked, they may well give
contradictory answers

• Human experts can be expensive & have full diaries

• Combinatorial explosion. To deal with absolutely every

eventuality, rules must be continually added to cover for every
possible situation, no matter how unlikely.

• Time - some expert systems contain thousands of rules, to deal

with something that may be simple all these rules must be
tested, along with conflict resolution

• Expert systems just deal with one aspect of intelligence

 Machine Learning
• Important though not to see such systems as merely
programmed decision making mechanisms that will
always perform as we expect
• It is possible to operate them in this way
• It is also possible to enable them to learn as they draw
conclusions and experience their domain
• If a system draws a number of conclusions then the
rules which result in the ‘winning’ conclusion can be
‘rewarded’ by making them more likely to be part of the
overall conclusion next time around.
• If a fired rule results in a conclusion which is not chosen
then it is less likely to fire again. Success is rewarded
and failure is punished - Bucket Brigade
 Fuzzy Logic & Fuzzy Rules
• With the expert systems we have considered thus far it
has been assumed that either a condition exists or it
doesn’t. This is Logic. A fact is either true or false

• It is found useful in certain circumstances for conclusions

to be partially true - for a confidence percentage to be
applied to results

• If someone is having a shower they want the water to be

warm. The water is not simply hot or cold, it is warm

• Fuzzy logic provides a basis for this.

 Example
• Assume for shower water to be completely cold it will be at
a temperature of 0 deg C - to be completely hot it will be at
a temperature of 50 deg C

• If the water is 65% hot – it is warm, but has some way to

go to be hot. If it is 12% hot then it is pretty cold

• Using fuzzy logic this does not necessarily mean that the
actual measured temperature would be 65% of 50, i.e. 32.5

• Fuzzy logic is more directed to a human concept of the

temperature – it is a form of artificial intelligence

• We can draw up a relationship between the actual

temperature and the percentage value we will assign it,
 Fuzzification

• A real-world value must be made fuzzy –

e.g. the water temperature is measured
and then fuzzified. A temperature of 20 oC
might become a fuzzy value of 45%
• The relationship between the actual value
and fuzzy value needs to be defined – this
could be done through graphical means or
a look up table
• The actual fuzzification routine depends
entirely on the application.
 Fuzzy Rules
• Once a value has been fuzzified it is passed to the rules for
evaluation.
• Fuzzy rules are the same as before, E.g. IF (water is hot) THEN (turn
water heater on)
• The condition part and conclusion part will be a percentage value as
they are only partially true. The water heater will not be turned on or
off, but will be turned on to a certain extent
• It may be that a rule has several conditions that need to be satisfied
• E.g. IF (water temperature is hot AND energy tariff is high) THEN
(turn water heater on)
• Each of the conditions has a percentage assigned to it. Where the
AND term appears so the MINIMUM percentage value of the
conditions is carried forward. Where the OR term appears so the
MAXIMUM percentage value of the conditions is carried forward.
• E.g. after fuzzification the water temperature has been assigned a
percentage 62% AND the energy tariff has been assigned a
percentage 48%. The value carried forward will be 48%.
• Normally a number of different rules will fire. Each of the rules will
result in a different value taken forward - these values must be
aggregated to provide a single value that actually means something
 Defuzzification
• Simplest is to average the values.
• We have three rules R1, R2 and R3 which have produced
the resultant percentage values. R1 – 23%, R2 – 81% and
R3 – 49% - the average value would be the three
percentages added together divided by three, i.e. 51%.
• Some rules may be more important than others - Centre
of Gravity (COG) method.
• If Rule R1 is more important than the others, give it a
weighting of 5, R2 a weighting of 2 and R3 a weighting of
3. We multiply 23 by 5, 81 by 2 and 49 by 3, the result of
which is 424, which, when we divide it by the sum=10,
gives us a defuzzified value of 42.4%. It is lower than the
unweighted calculation as emphasis was placed on R1
 Problem Solving
• We may need an AI system to solve problems for us.
One example of this exists in a SatNav. We know where
our start point is and we know where we wish to get to,
but we don’t know how to get there.
• Many solutions exist. We usually have further
requirements, we wish to know the quickest route or the
shortest one. This problem is something that AI can be
very good at solving - quickly.
• Assume we wish to travel from Reading to Newcastle.
Many possible routes – we could travel from Reading to
Oxford or Reading to London. Both routes have costs
associated with them - the time the route would take,
the petrol used and so on. From Oxford we could travel
to Banbury etc. Until Newcastle was reached.
• Assume we limit the number of possible towns to be
considered from Reading to Newcastle and we only visit
a town once - there are a number of ways that an AI
system could search for the best solution.
 Searching
• To decide the best solution to our travel problem we must
consider all possibilities

• In our example with Reading as a start point then we could

search for the best route by firstly looking at all the
possible towns to travel to from Reading – Oxford and
London included. From each of those towns we could then
look at all choices - cost

• We would eventually arrive at Newcastle via a number of

different routes. As we know the total cost of each route
then we can decide which is best in terms of distance, time
or whatever
 Methods

• Depth first search

• Breadth first search
• Depth limited search
• Bidirectional search
• Heuristic search
 Example

• In May 1997 the IBM computer ‘Deep Blue’ beat the

erstwhile human world chess champion, Gary
Kasparov, over a 6 match series.
• The computer was capable of searching and
analyzing 200 million positions every second
• Kasparov said: “There were many, many discoveries
in this match, and one of them was that sometimes
the computer plays very, very human moves. It
deeply understands positional factors”.
 Frames
• A frame represents everyday knowledge about an
entity
• It is a computer file, with a number of pieces of
information stored in slots in the file. Each of
those slots is a sub-frame with further levels of
information.
• If we have a frame based artificial intelligence
system used to describe a house – the initial
frame is the house
• Within the house are slots, dining room, kitchen,
lounge. Each slot is a frame itself. A kitchen
frame contains slots, refrigerator, cooker, sink
etc. These slots are frames with their own slots.
 Data Mining
• Extracting knowledge from data is called data mining.
• AI systems are well suited - they store enormous quantities
of data and can draw out all sorts of relationships
• The amount of data in the world doubles each year – in a
ten year span (e.g. 2002 to 2012) there will be a 1,000
times increase in data!
• New data is usually not well understood and meanings are
not readily drawn out. E.g. The human genome project has
opened up the complexities of DNA - we can look at the
functioning of brains and make sense of it based on the data
• There are new business opportunities, new medical
techniques and a more in depth understanding of the
scientific world
 Correlations
• To discover similarities, links and relationships between data and to
predict outcomes
• Example - supermarket shopping. Such shopping is a regular exercise.
Approximately 100 different types of produce available. Every time a
person uses the supermarket, data is obtained on their purchases
• Links can be drawn as to what a person buys and how often they buy
it OR which people buy certain products and when they buy them
• We want a prediction - next Thursday a person will enter the
supermarket, they will buy certain products – if they are available,
the person will buy other products based on our predictions. Over
1,000 people it may well be sufficiently accurate for a significant profit
• Correlation - Look at a person’s purchases of milk and cheese. Over a
one year period this can be analyzed to see how the two pieces of
data are related to each other.
• A number of statistical tools are available for this, Principal
Component Analysis, which detects the main links between pieces of
data, e.g. for one person the purchasing of shoe polish may be closely
linked to buying pickles.
 Decision Trees

• To reduce the complexity of problems to make a database

easier to analyze - decision trees.
• The data base is chopped up into manageable portions
which makes it easy to follow a path through the tree.
• For the supermarket example, we could only consider female
purchasers. A user specified branch - only data associated
with female purchasers need be considered
• We can input requirements, the branches of which could be
discovered as a part of the analysis. E.g. only those who
spend more than £60, regularly purchase soup and buy
fresh vegetables.
• Rather than dealing with a large number of people, say
50,000, we may only need to consider 1,000, which will
reduce the time taken and improve the accuracy of results
 Concluding Comments
• Classical AI techniques are based on trying to get
machines to copy humans in tasks that when humans do
those tasks, we deem them to be intelligent acts
• Discussion ranged from the way we store information
(frames) to how we reason and decide (expert systems)
• Motivation - advantages of artificial intelligence when
compared with human intelligence – to replace humans!
• Speed of processing, accuracy of mathematical
calculations, extent of memory, relating complex data
and the ability to function 24/7
• Intelligence is a controversial topic
• When we consider machines being intelligent this raises
the stakes
• How does machine intelligence compare with human
intelligence? Can a machine be alive? Next we look at
philosophical issues that underpin the subject
 Next

• Philosophy of New AI