Transcript Lecture 5

Neuroevolution and
Developmental Encoding
Evolutionary Comptation
Dr. Kenneth Stanley
September 11, 2006
Main Idea
• Natural selection can work on computers
– Selection: Picking the best parents
– Variation: Mutation and Mating
Start with some really bad individuals
Some are always better than others
Survival of the fittest leads to improvement
Progress occurs over generations
Survival of the Roundest
Gen 1
Select as parents
Gen 2
Select as parents
Gen 3
Several Versions of EC
Genetic Algorithms (Holland 1960s)
Evolution Strategies (Rechenberg 1965)
Evolution Programming (Fogel 1966)
Genetic Programming? (Smith 1980,Koza 1982)
The process is more important than the name
Major Concepts
Genotype and Phenotype
Representation / mapping
Evaluation and fitness
Steady state
Premature Convergence
Genotype and Phenotype
• Genotype means the code (e.g. DNA) used to
the describe an organism, i.e. the “blueprint”
• Phenotype is the organism’s actual realization
f ( x)  3x 2  7 x  10
Representation and Mapping
• The genotype is a representation of the
phenotype; how to represent information is
a profound and deep issue
• The process of creating the phenotype
from the genotype is called the genotype
to phenotype mapping
• Mapping can happen in many ways
Evaluation and Fitness
• The phenotype is evaluated, not the
• The performance of the phenotype during
evaluation is its fitness
• Fitness tells us which genotypes are better
than others
• Most GAs proceed in generations:
– A whole population is evaluated one at a time
– That is the current generation
– They then are replaced en masse by their
– The replacements form the next generation
– And so on…
Steady State Evolution
• Not all EC is generational
• It is possible to replace only one individual
at a time, i.e. steady state evolution
• Common in Evolution Strategies (ES)
• Also called real-time or online evolution
• Another twist: Phenotypes can be
evaluated simultaneously and
• Selection means deciding who should be a
parent and who should not
• Selection is usually based on fitness
• Methods of selection (see Mitchell p.166)
– Roulette Wheel (probability based on fitness)
– Truncation (random among top n%)
– Rank selection (use rank instead of fitness)
– Elitism (champs get to have clones)
• Mutation means changing the genotype
• Can vary from strong (every gene
mutates) to weak (only one gene mutates)
• May mean adding a new gene entirely
• Mutation prevents fixation
• Mutation is a source of diversity and
• Combining one or more genomes
• Many ways to implement crossover:
– Singlepoint
– Multipoint (Uniform)
– Multipoint average (Linear)
• How important is crossover?
• What is it for?
Premature Convergence
• When a single genotype dominates the
population, it is converged
• Convergence is premature if a suitable
solution has not yet been found
• Premature convergence is a significant
concern in EC
• Hence the need to maintain diversity
A population can be divided into species
Can prevent incompatibles from mating
Can protect innovative concepts in niches
Maintains diversity
Many methods
– Islands
– Fitness sharing
– Crowding
Natural Evolution is not Just
What is the optimum?
What is the space being searched?
What are the dimensions?
Herb Simon (1958): “Satisficing”
Is evolution even just a satisficer?
Evolution satisfices and complexifies
Next Class:
Theoretical Issues in EC
• The Schema Theorem
• No Free Lunch
Mitchell pp. 117-38, and ch.5 (pp. 170-177)
No Free Lunch Theorems for Optimization
by Wolpert and Macready (1996)