statistical mechanics for g as
TRANSCRIPT
Statistical Mechanics for GAsA gentle introduction
Presented by :
Yann SEMET
Universite de Technologie de Compiegne
Overview
Motivations and general idea
Definitions
Selection analysis (pb. independent)
Problem specific analysis : mutation, crossover
Results on OneMax
Beyond simple problems
2 documents
An analysis of Genetic Algorithms Using Statistical Mechanics
Prugel-Bennett, Shapiro. 1994
Modeling the dynamics of Gas using Statistical Mechanics
Rattray. 1996
Motivations
Markov chains :Exact model
Gets intractable with size
Statistical mechanicsProbabilistic model
More compact
Macroscopic description
Macroscopics
Cumulants
Mean correlation
Evaluates the evolution of fitness distribution
Modelling the dynamics
Each genetic operator :A set of difference equations (on macroscopics)
Iteration
Non trivial terms : Maximum entropy ansatz
Finite population effectsA finite sample from an infinite population
Selection
Infinite population again
Definitions
Genotype, phenotype and fitness :
Fitness distribution :
Cumulants
Definition :
First two :
Cumulants (cont.)
Infinite population :
Finite sample corrections :
Gram-Charlier expansion
A convenient approximation
Correlation
A measure of genotype similarity
Mean value :
Best population member
Our goal after all
Modelling selection
Problem independent
A general scheme :
2 stages :Random sampling from infinite population
Generating a new infinite one
Selection (cont.)
Generating the cumulants :
Selection (cont.)
Expansion :
Finally :
Correlation after selection
2 terms :
Duplication
Natural change (problem specific)
Final approximation :
Selection schemes
Particular schemes :Boltzmann
Truncation
Ranking
Tournament
Tackling problems
Problem specific operators
A convenient class of problems : Functions of an additive genotype
Cumulants :
Mutation (1)
Mutation (2)
Cumulants :
Notice non trivial terms
Correlation :
Crossover
A generalized form of uniform crossover :
Cumulants :
Mean correlation unchanged
Maximum entropy ansatz
Calculate terms non trivially related to known macroscopics
Assumptions on allele distribution
2 constraintsMean phenotype
Correlation
Results
Onemax problem
2nd class of problems
Fitness=stochastic function of phenotype
Test problem : perceptron with binary weights
Competent model :Size population accurately to remove noise
Size training batch consequently
A NP hard problem
Storing random pattern in a binary perceptron
Insight gained
Half failure :Technical difficulties
Inconsistencies
Incomplete model
Extensions of the model
Two tests :1 simple diploid GA
1 temporally varying fitness
Successful description under :Bit-simulated crossover
Extra constraints for MEA
Summary
Motivation of macroscopic models
Cumulants
Mean correlation
Dynamics modeling
Simple problems
Extensions
Conclusions
StrengthsCompactness and accuracy
Finite population effects
WeaknessesLimitations of MAE
NP hard problem inaccurately modeled
Technical limitation
Fundamental limitations ?
Punctuated equilibria