what is neutral?
DESCRIPTION
What is Neutral?. Neutral Changes and Resiliency Terence Soule Department of Computer Science University of Idaho. The experiments. Gene/exon selection Introns and exon selection Effects of operators. +. +. 1.0. 0.5. 1.0. Experiment 1. Tree based, generational GP Functions {+} - PowerPoint PPT PresentationTRANSCRIPT
What is Neutral?Neutral Changes and
Resiliency
Terence SouleDepartment of Computer Science
University of Idaho
The experiments
Gene/exon selectionIntrons and exon selectionEffects of operators
Experiment 1Tree based, generational GPFunctions {+}Terminals/Genes {0.5, 1.0}Fitness: difference from 10
Both terminals are exons. Is one selected?
+
+
0.5
1.0
1.0
Gene/Exon Choice
0
0.25
0.5
0 20 40 60 80 100Generation
Per
cent
of c
ode
by
gene
1.00.5+
Average Fitness
2.4
2.6
2.8
3
0 20 40 60 80 100
Generation
Av
erg
e F
itn
es
s
Average fitness improves – after crossover.
Resiliency
A measure of expected fitness change as a function of genotype change.Resilient individuals are less likely to change fitness or have a smaller average fitness change in response to genotype changes (crossover and mutation).Similar to the idea of effective fitness, but more general.
Experiment 2Tree based, generational GPFunctions {+}Terminals/Genes {0, 1, 4}Fitness: difference from 40
Now there are two exons and an intron. What is selected?
Number of Genes
020406080
100120140160180
0 500 1000 1500 2000
Num
ber
of
Generation
0s1s4s
Resiliency
-6
-5
-4
-3
-2
-1
0
1
0 250 500 750
Fit
ness
Cha
nge
CrossoverMutation
Ratio of 1s to 4s
Results – Experiment 2Changes don’t affect current fitness – Are they Neutral?Changes affect expected fitness of the next generation – increase (average) resiliency
Experiment 3Variable length, linear encoding, generationalGenes {0, 1, 4}Sample individual: 010041014Fitness: difference of sum of genes from 54
Experiment 3 - Crossover
Proportional crossover – select two random points per parent.Constant crossover – length of crossed region is:
2 50% of the time4 25% of the time8 12.5% of the time…
00 104 04
440 01011 0401
00 01011 04
440 104 0401
Genes – Constant Crossover
0
50
100
150
200
0 200 400 600 800 1000Generation
Nu
mb
er
of
0s
1s
4s
Genes – Proportional Crossover
9
10
11
12
13
0 200 400 600 800 1000Generation
Nu
mb
er
of
0s
1s
4s
Mutation – Constant Crossover
Probability P of changing a gene to another value: 1 to 0, etc.More genes (including 0s) greater chance of mutations.
Growth – constant crossover
ConclusionsMany ‘neutral’ changes can be explained in terms of resiliency
1.0 two 0.5s (selecting exons)4s four 1s and four 1s one 4sIncreasing 0s (increasing introns)
Operator choice significantly affects these changes
Proportional versus constant crossoverMutations
Per node versus per individual rates are significant.
DiscussionTypes of changes
1st order – affect fitness2nd order – affect expected fitness of offspring (resiliency)3rd order? - affect expected fitness of Nth generation? Affect ability to respond to ‘environmental’ changes?
Any consistent pattern of change has an evolutionary explanation(?)It’s possible to predict some changes by using the idea of resiliency.Do these changes affect search?
Thank You
Questions?
Bibliography“Exons and Code Growth in GP” Genetic Programming 5th European Conference, EuroGP-2002, Springer LNCS2278, 2002 .“Solution Stability in Evolutionary Computation” Proceedings of the 17th International Symposium on Computer and Information Sciences, CRC Press, 2002.“Operator Choice and the Evolution of Robust Solutions” Genetic Programming Theory and Practice, Kluwer, 2003.