Reading
CS 420/594: Read Flake, ch. 22 (Neural Networks and Learning)
CS 594: Read Bar-Yam, sec. 2.3 (Feedforward Networks)

Why Does the GA Work?
The Schema Theorem

Schemata
A schema is a description of certain patterns of bits in a genetic string

The Fitness of Schemata
The schemata are the building blocks of solutions
We would like to know the average fitness of all possible strings belonging to a schema
We cannot, but the strings in a population that belong to a schema give an estimate of the fitness of that schema
Each string in a population is giving information about all the schemata to which it belongs (implicit parallelism)

Effect of Selection
Exponential Growth
We have discovered:
m(S, t+1) = m(S, t)
× f(S) / fav
Suppose f(S) = fav (1 + c)
Then m(S, t) = m(S, 0) (1 + c)t
That is, exponential growth in above-average schemata

Effect of Crossover
Let l = length of genetic strings
Let d(S) = defining length of schema S
Probability {crossover destroys S}:
pd = d(S) / (l Ð 1)
Let pc = probability of crossover
Probability schema survives:

Selection & Crossover Together
Effect of Mutation
Let pm = probability of mutation
So 1 Ð pm = probability an allele survives
Let o(S) = number of fixed positions in S
The probability they all survive is
(1 Ð pm)o(S)
If pm << 1, (1 Ð pm)o(S) Å 1 Ð o(S) pm

Schema Theorem:
ÒFundamental Theorem of GAsÓ
The Bandit Problem
Two-armed bandit:
random payoffs with (unknown) means m1, m2  and variances s1, s2
optimal strategy: allocate exponentially greater number of trials to apparently better lever
k-armed bandit: similar analysis applies
Analogous to allocation of population to schemata
Suggests GA may allocate trials optimally

GoldbergÕs Analysis of Competent & Efficient GAs
Paradox of GAs
Individually uninteresting operators:
selection, recombination, mutation
Selection + mutation Þ continual improvement
Selection + recombination Þ innovation
generation vs.evaluation
Fundamental intuition of GAs: the three work well together

Race Between Selection & Innovation: Takeover Time
Takeover time t* = average time for most fit to take over population
Transaction selection: top 1/s replaced by s copies
s quantifies selective pressure
Estimate t* Å ln n / ln s

Innovation Time
Innovation time ti = average time to get a better individual through crossover & mutation
Let pi = probability a single crossover produces a better individual
Number of individuals undergoing crossover = pc n
Probability of improvement = pi pc n
Estimate: ti Å 1 / (pc pi n)

Steady State Innovation
Bad: t* < ti
because once you have takeover, crossover does no good
Good: ti < t*
because each time a better individual is produced, the t* clock resets
steady state innovation
Innovation number:

Feasible Region
Other Algorithms Inspired by Genetics and Evolution
Evolutionary Programming
natural representation, no crossover,time-varying continuous mutation
Evolutionary Strategies
similar, but with a kind of recombination
Genetic Programming
like GA, but program trees instead of strings
Classifier Systems
GA + rules + bids/payments
and many variants & combinationsÉ

Additional Bibliography
Goldberg, D.E.  The Design of Innovation: Lessons from and for Competent Genetic Algorithms.  Kluwer, 2002.
Milner, R.  The Encyclopedia of Evolution.  Facts on File, 1990.