Genetic Algorithms

26-May-13

Evolution

Heres a very oversimplified description of how evolution works in biology Organisms (animals or plants) produce a number of offspring which are almost, but not entirely, like themselves

Variation may be due to mutation (random changes) Variation may be due to sexual reproduction (offspring have some characteristics from each parent)

Some of these offspring may survive to produce offspring of their ownsome wont

The better adapted offspring are more likely to survive Over time, later generations become better and better adapted

Genetic algorithms use this same process to evolve better programs

Genotypes and phenotypes

Genes are the basic instructions for building an organism A chromosome is a sequence of genes Biologists distinguish between an organisms genotype (the genes and chromosomes) and its phenotype (what the organism actually is like)

Example: You might have genes to be tall, but never grow to be tall for other reasons (such as poor diet)

Similarly, genes may describe a possible solution to a problem, without actually being the solution
3

The basic genetic algorithm

Start with a large population of randomly generated attempted solutions to a problem Repeatedly do the following: Evaluate each of the attempted solutions Keep a subset of these solutions (the best ones) Use these solutions to generate a new population Quit when you have a satisfactory solution (or you run out of time)

A really simple example

Suppose your organisms are 32-bit computer words You want a string in which all the bits are ones Heres how you can do it:

Create 100 randomly generated computer words Repeatedly do the following: Count the 1 bits in each word Exit if any of the words have all 32 bits set to 1 Keep the ten words that have the most 1s (discard the rest) From each word, generate 9 new words as follows: Pick a random bit in the word and toggle (change) it

Note that this procedure does not guarantee that the next generation will have more 1 bits, but its likely
5

A more realistic example, part I

Suppose you have a large number of (x, y) data points

For example, (1.0, 4.1), (3.1, 9.5), (-5.2, 8.6), ...

You would like to fit a polynomial (of up to degree 5) through these data points

That is, you want a formula y = ax5 + bx4 + cx3 + dx2 +ex + f that gives you a reasonably good fit to the actual data Heres the usual way to compute goodness of fit:

Compute the sum of (actual y predicted y)2 for all the data points The lowest sum represents the best fit

There are some standard curve fitting techniques, but lets assume you dont know about them You can use a genetic algorithm to find a pretty good solution
6

A more realistic example, part II

Your formula is y = ax5 + bx4 + cx3 + dx2 +ex + f Your genes are a, b, c, d, e, and f Your chromosome is the array [a, b, c, d, e, f] Your evaluation function for one array is:

For every actual data point (x, y), (Im using red to mean actual data)

Compute = ax5 + bx4 + cx3 + dx2 +ex + f Find the sum of (y )2 over all x The sum is your measure of badness (larger numbers are worse)
= 0x5 + 0x4 + 0x3 + 2x2 +3x + 5 is 2 + 3 + 5 = 10 when x is 1 = 0x5 + 0x4 + 0x3 + 2x2 +3x + 5 is 8 + 6 + 5 = 19 when x is 2 (12 10)2 + (22 19)2 = 22 + 32 = 13 If these are the only two data points, the badness of [0, 0, 0, 2, 3, 5] is 13
7

Example: For [0, 0, 0, 2, 3, 5] and the data points (1, 12) and (2, 22):

A more realistic example, part III

Your algorithm might be as follows:

Create 100 six-element arrays of random numbers Repeat 500 times (or any other number): For each of the 100 arrays, compute its badness (using all data points) Keep the ten best arrays (discard the other 90) From each array you keep, generate nine new arrays as follows: Pick a random element of the six Pick a random floating-point number between 0.0 and 2.0 Multiply the random element of the array by the random floating-point number After all 500 trials, pick the best array as your final answer

Asexual vs. sexual reproduction

In the examples so far,

Each organism (or solution) had only one parent Reproduction was asexual (without sex) The only way to introduce variation was through mutation (random changes) Each organism (or solution) has two parents Assuming that each organism has just one chromosome, new offspring are produced by forming a new chromosome from parts of the chromosomes of each parent

In sexual reproduction,

The really simple example again

Suppose your organisms are 32-bit computer words, and you want a string in which all the bits are ones Heres how you can do it:

Create 100 randomly generated computer words Repeatedly do the following: Count the 1 bits in each word Exit if any of the words have all 32 bits set to 1 Keep the ten words that have the most 1s (discard the rest) From each word, generate 9 new words as follows: Choose one of the other words Take the first half of this word and combine it with the second half of the other word

10

The example continued

Half from one, half from the other: 0110 1001 0100 1110 1010 1101 1011 0101 1101 0100 0101 1010 1011 0100 1010 0101 0110 1001 0100 1110 1011 0100 1010 0101 Or we might choose genes (bits) randomly: 0110 1001 0100 1110 1010 1101 1011 0101 1101 0100 0101 1010 1011 0100 1010 0101 0100 0101 0100 1010 1010 1100 1011 0101

Or we might consider a gene to be a larger unit: 0110 1001 0100 1110 1010 1101 1011 0101 1101 0100 0101 1010 1011 0100 1010 0101 1101 1001 0101 1010 1010 1101 1010 0101
11

Comparison of simple examples

In the simple example (trying to get all 1s):

The sexual (two-parent, no mutation) approach, if it succeeds, is likely to succeed much faster

Because up to half of the bits change each time, not just one bit By pure bad luck, maybe none of the first (randomly generated) words have (say) bit 17 set to 1 Then there is no way a 1 could ever occur in this position Another problem is lack of genetic diversity Maybe some of the first generation did have bit 17 set to 1, but none of them were selected for the second generation

However, with no mutation, it may not succeed at all

The best technique in general turns out to be sexual reproduction with a small probability of mutation
12

Curve fitting with sexual reproduction

Your formula is y = ax5 + bx4 + cx3 + dx2 +ex + f Your genes are a, b, c, d, e, and f Your chromosome is the array [a, b, c, d, e, f] Whats the best way to combine two chromosomes into one?

You could choose the first half of one and the second half of the other: [a, b, c, d, e, f] You could choose genes randomly: [a, b, c, d, e, f] You could compute gene averages: [(a+a)/2, (b+b)/2, (c+c)/2, (d+d)/2, (e+e)/2,(f+f)/2]

I suspect this last may be the best, though I dont know of a good biological analogy for it
13

Directed evolution

Notice that, in the previous examples, we formed the child organisms randomly

We did not try to choose the best genes from each parent This is how natural (biological) evolution works

Biological evolution is not directedthere is no goal

Genetic algorithms use biology as inspiration, not as a set of rules to be slavishly followed

For trying to get a word of all 1s, there is an obvious measure of a good gene

But thats mostly because its a silly example Its much harder to detect a good gene in the curve-fitting problem, harder still in almost any real use of a genetic algorithm
14

Probabilistic matching

In previous examples, we chose the N best organisms as parents for the next generation A more common approach is to choose parents randomly, based on their measure of goodness

Thus, an organism that is twice as good as another is likely to have twice as many offspring The best organisms will contribute the most to the next generation Since every organism has some chance of being a parent, there is somewhat less loss of genetic diversity

This has a couple of advantages:

15

Genetic programming

A string of bits could represent a program If you want a program to do something, you might try to evolve one As a concrete example, suppose you want a program to help you choose stocks in the stock market

There is a huge amount of data, going back many years What data has the most predictive value? Whats the best way to combine this data?

A genetic program is possible in theory, but it might take millions of years to evolve into something useful

How can we improve this?

Shrinking the search space

There are just too many possible bit patterns!

99.9999% of these dont even represent valid programs

An incredible improvement would result if we could somehow restrict the search space to only valid (even if nonsensical) programs

We can do this!

Programs, as you should know by now, can be represented as trees

Internal nodes are operators: +, *, if, while, ... Leaves are values: 2.71818, "AAPL", ...
17

Programs as trees

Given a program represented as a tree, we can mutate it by changing one of its operators (or one of its values), or by adding or removing nodes Given two trees, we can form a new tree by taking parts of its two parents The next big problem: How do we evaluate program trees that are (initially) nothing at all like what we want? I realize this is all very vagueI just wanted to give you the general idea
18

Concluding remarks

Genetic algorithms are

Fun! They are enjoyable to program and to work with

This is probably why they are a subject of active research

Mind-bogglingly slowyou dont want to use them if you have any alternatives Good for a very few types of problems

Genetic algorithms can sometimes come up with a solution when you can see no other way of tackling the problem

19

The End

20