Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Phenotype to Genotype Matching and Epigeneticsin Evolutionary Algorithms
William Maroney
Australian National University
u5612989
May 21, 2017
Supervisors: Tom Gedeon and Bob McKay
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Overview
Evolutionary computing background
Research motivation
Possible epigenetic models for the genetic algorithm
Experimental design and findings
Future work
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
The Genetic Algorithm
An optimisation technique
Useful for large, complex,or undefined search spaces
Uses theory of evolution:survival of the fittest
Requires fitness function:how “good” is a givencandidate solution?
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Interactive Evolutionary Computing (IEC)
Evaluation of the fitnessfunction requires humaninput
Example: how much doyou like this? Rate it.
Computing power doesn’thelp us . . .
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Research motivation
Can we improve performance of the genetic algorithm?
Can we reduce the impact of humans in IEC?
i.e. find good solutions faster
The genetic algorithm works pretty well with a simplisticmodel of the theory of evolution. Does a moreaccurate/comprehensive representation improve things?
Consider epigenetics?
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Possible Epigenetic Models
Traditional genetic algorithm: assign whole fitness valuef (g) := f (m∗(g))
Epigenetics (exact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p | g)× 1
||m∗(g)−p|| (all evaluated p, time ≤ τ)
Epigenetics (inexact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p′ | g)× 1
||p−p′|| ×1
||m∗(g)−p|| (closet p′ to p)
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Possible Epigenetic Models
Traditional genetic algorithm: assign whole fitness valuef (g) := f (m∗(g))
Epigenetics (exact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p | g)× 1
||m∗(g)−p|| (all evaluated p, time ≤ τ)
Epigenetics (inexact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p′ | g)× 1
||p−p′|| ×1
||m∗(g)−p|| (closet p′ to p)
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Possible Epigenetic Models
Traditional genetic algorithm: assign whole fitness valuef (g) := f (m∗(g))
Epigenetics (exact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p | g)× 1
||m∗(g)−p|| (all evaluated p, time ≤ τ)
Epigenetics (inexact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p′ | g)× 1
||p−p′|| ×1
||m∗(g)−p|| (closet p′ to p)
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Possible Epigenetic Models
Traditional genetic algorithm: assign whole fitness valuef (g) := f (m∗(g))
Epigenetics (exact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p | g)× 1
||m∗(g)−p|| (all evaluated p, time ≤ τ)
Epigenetics (inexact P matches): proportionally infer fitnessf τ (g) :=
∑p f (p)× p(p′ | g)× 1
||p−p′|| ×1
||m∗(g)−p|| (closet p′ to p)
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Experimental Set-up
Multiple test subjects
Each test subject evaluated all three models
GA model, epigenetic model (exact and inexact matches)Same number of generations for each modelModel order assigned to avoid possible bias
Consistent hyper-parameters (i.e. only compare models)
Need to avoid user fatigue ⇒ limited fitness evaluations
Unavoidable challenge with IEC
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Performance Measures
Hypothesis: epigenetic fitness inference increases convergence
Hypothesis: epigenetic fitness inference affects convergence
Ratio of positive artworks to all artworks over time
Identify affect on fitness values - Mean Absolute Error (MAE)
Compare genetic model fitness function to each epigeneticmodel fitness functionConsider MAE per generation
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Results
Experiments are suggestive, not conclusive
Unclear that epigenetic fitness inference increases convergence
However, fitness values are affected
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Experimental Results
All models produce non-random results (i.e. consistentlybetter than“indifferent” rating)
Hard to infer if any model is clearly superior
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Experimental Results
Some effect on fitness values ⇒ altered selection
Actual impact and importance still unclear
Phenotype to Genotype Matching and Epigenetics in Evolutionary Algorithms
Future Work
Further investigation into experimental implications
Increase scale of experiments performed in this work
Apply epigenetic models to other problems (IEC, and EC)
More investigation into epigenetic fitness functionhyper-parameters (i.e. scale and normalisation factors)