Scribd
Upload
82 views29 pages

Niching and Speciation

This document discusses techniques called niching and speciation that are useful for maintaining diversity in evolutionary algorithm populations. Niching methods like fitness sharing and crowding allow populations to find multiple optimal solutions by preventing the population from converging to a single area. Speciation methods further improve results by restricting mating to only similar individuals, mimicking how species evolve separately in nature. Simulation results showed speciation methods outperform niching alone on test functions with multiple peaks.

Uploaded by

Chris Axl
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
82 views29 pages

Niching and Speciation

This document discusses techniques called niching and speciation that are useful for maintaining diversity in evolutionary algorithm populations. Niching methods like fitness sharing and crowding allow populations to find multiple optimal solutions by preventing the population from converging to a single area. Speciation methods further improve results by restricting mating to only similar individuals, mimicking how species evolve separately in nature. Simulation results showed speciation methods outperform niching alone on test functions with multiple peaks.

Uploaded by

Chris Axl
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 29

Niching and Speciation

Fernando Lobo

University of Algarve
Introduction

•  We will look at techniques that are useful to maintain


diversity in the population.

•  Especially useful to find multiple (different) solutions in


multimodal problems. Also useful in multi-objective
optimization, and simulation of complex and adaptive
systems.

•  Methods inspired in Nature:


–  Niching
–  Speciation
Multimodal problems

•  In problems with multiple optima, a standard EA


eventually focuses its attention on a single optima.
–  Even if there’s many equally good ones.

•  Why is that?
–  Due to chance variation alone, the population will end
up being in the vicinity of a single solution.
Methods

•  We will be looking at two methods inspired in Nature:


–  Niching
–  Speciation
Niching

•  Niching methods allow EAs to maintain a population of


diverse solutions through time.

•  EAs with niching are capable of obtaining multiple optimal


solutions in a single population.

•  Motivation from Nature:


–  natural evolutionary processes maintain a variety of
species, each occupying its own ecological niche.
Niching

•  Two major methods:


–  Fitness sharing
–  Crowding
Fitness sharing

•  Proposed by Goldberg and Richardson (1987).

•  Idea:
–  Resources in Nature are limited.
–  Similar individuals (those occupying the same niche)
must share those resources (fitness) with each other.
Fitness sharing (cont.)
•  Sharing works by degrading an individual’s fitness by an
amount related to the number of similar individuals
contained in the population.

•  The shared fitness f´ is equal to the original fitness f,


divided by its niche count.

•  An individual’s niche count is the sum of sharing function


(sh) values between itself and each individual in the
population.
Fitness sharing (cont.)
•  The sharing function is a function of the distance between
two individuals:
–  returns 1 if the two individuals are identical
–  returns 0 if the the distance is above a certain threshold
( , a user defined parameter).
–  returns something in between 0 and 1 depending on the
level of similarity.
Fitness sharing (cont.)
•  A common share function is:

•  α is a constant (typically set to 1).

•  distance d can be computed at the genotype level


(Hamming distance for bit strings) or at the phenotype
level.

•  should be set to allow discrimination between


desired peaks. (in practice this may be unknown a priori.)
Crowding

•  First proposed by De Jong (1975).

•  Idea: new individuals replace similar individuals in the


population.

•  Also use a distance function to measure similarity.

•  Unlike sharing, crowding methods do not allocate


individuals proportional to each niche (peak) fitness.

•  By replacing similar individuals, crowding promotes pre-


existing diversity of a population.
Crowding (cont.)

•  There’s various ways to implement this mechanism.

•  Let’s look at the original method proposed by De Jong.

•  The algorithm is a steady-state EA: only a fraction of the


population reproduces and dies each generation.
De Jong’s original crowding model

•  A proportion GG (generation gap) of the population is


chosen via fitness-proportionate selection to undergo
crossover and mutation.

•  For each new individual, take a a random sample of CF


(crowding factor) individuals from the population. The
new individual replaces the one that is most similar to it in
the random sample.

•  (GG is typically around 0.1, and CF around 2-5)


Simulation results: Sharing vs Crowding

•  Results taken from Deb and Goldberg (An investigation of


niche and species formation in genetic function
optimization, ICGA 1989).

•  Two test functions: F1 and F2


–  F1: equal peaks
–  F2: non-equak peaks
F1: equal peaks

F1(x) =
F2: unequal peaks

F2(x) =
GA Parameters

•  set according to Deb and Goldberg’s


recommendations (see paper for details).
F1 results
F2 results
Results (cont.)

•  Crowding is unable to maintain the multiple peaks


(maintains only 2 for F1 and only 1 for F2).

•  On peaks with unequal value, genotypic sharing is not able


to maintain the lower peaks.
Restricted Tournament Selection (RTS)

•  RTS is a crowding-like method and works very well. It


was proposed by Harik (1995).

•  RTS is a modification of tournament selection that restricts


an individual from competing with others that are very
different from it.
RTS works as follows

•  Select two solutions at random, A and B, from the


population and perform crossover and mutation, resulting
in two new solutions, A´ and B´.

•  For each new solution (A´ and B´), scan w individuals


(randomly chosen) from the population and pick the one
that is most similar to it. Call them A´´ and B´´.

•  A´ competes with A´´. If A´ is better, then it replaces A´´


in the population.
–  The algorithm is steady-state (solutions are introduced
in the population incrementally).
RTS

•  RTS performs much better than the standard crowding


mechanism.

•  Also much simpler to use than fitness sharing.

•  Harik used w as 4 times the number of desired peaks to be


found, and obtained good results.
Speciation

•  Finding multiple solutions can be improved with


speciation methods.

•  A species is a collection of individuals which resemble


each other more closely than they resemble individuals of
another species.

•  Artificial species can be created in EAs by only allowing


mating between similar individuals.
–  In Nature, lions don’t mate with elephants!
Speciation (cont.)

•  Niching distributes individuals among multiple peaks.

•  But niching does not prevent an individual from one peak


(niche) to mate with an individual from another peak.
–  Such matings often result in so-called lethal solutions,
representing none of the peaks.

•  Can solve this problem with speciation methods.


Speciation methods

•  Deb and Goldberg proposed two mating restricting


schemes based on phenotypic and genotypic distances
between mating individuals.

•  Idea: if the distance is closer than a parameter , they


participate in the crossover operation. Otherwise another
individual is chosen at random as a candidate mate.
–  This process continues until a proper mate is found. If
no such member exists, then a random individual from
the population is chosen as mate.
Speciation methods

•  Niching with is implemented in the selecion


operator.

•  Mating restriction with is implemented in the


crossover operator.

•  Deb and Goldberg used

•  Results on F1 and F2 were improved.


F1 with and without mating restriction
Other speciation method

•  Using tag bits (Spears, 1994)

•  Tag bits are appended to every individual.

•  Each species correspond to a particular configuration of


those bits.

•  Mating is restricted to individuals containing the same set


of tag bits.

•  Spears allowed the tag bits to be mutated.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy