Niching and Speciation

Niching and Speciation 



Outline 

.Review of the last lecture 

.A motivating example of niching: co-evolution in 

classification tasks 

.Why niching 

.Different niching techniques 

.Sharing and crowding 

.Relationship between niching and speciation 

.Summary 



Different Niching Techniques 

.Can be divided roughly into two major categories 

.Sharing, also known as fitness sharing 

.Crowding 

.Other niching methods include sequential niching and parallel 

hillclimbing 



Fitness Sharing: Introduction 

.Fitness sharing transforms the raw fitness of an individual into 

shared fitness 

.It assumes that there is only limited and fixed “resource” 

available at each niche. Individuals in a niche must share them 

.Sharing is best explained from a multimodal function 

optimization perspective 

individual 

raw 

fitness 

How can we locate multiple peaks in one evolutionary process? 



Fitness Sharing: Implementation 

.Define a sharing radius sshare: Anything within this radius will be 

regarded to be similar to the individual and thus needs to share 

fitness 

an individual 

sshare 

The individual in the center needs to 

share fitness with all other in the circle 

.Define a similarity measure, i.e., distance: The shorter the 

distance between two individuals, the more similar they are 

.Define a sharing function 

a scaling constant 

1 . 

) 

a 

if d < sshare 

(d / sshare

sh(d) = 

0 otherwise 

distance 

.Define shared fitness 

m 

(i) =f(i) / Sj=1sh(dij)

fshareraw

where m is the population size 



Fitness Sharing: Extensions 

.Sharing can be done at genotypic or phenotypic level 

.Genotypic: Hamming distance 

.Phenotypic: Euclidean distance (Overlap in test case 

covering in classification) . 

The key issue is how to define 

the “distance” measure 

.Sharing radius sshare can be difficult to set, the same, fixed: It 

should be sufficiently small in order to discriminate between two 

neighboring peaks 

.Population size should be sufficiently large to locate all peaks 

.Population may not be able to converge to exact optima 

.Population may not be stable, i.e., may lose peaks located 

.Calculate shared fitness needs time 

scaling factor 

.Fitness sharing often needs raw fitness scaling. b 

m 

(i) =f(i) / Sj=1sh(dij)

Why? fshareraw



Why Fitness Scaling 

.Let fi(i),fi= f(i),mi 

m sh(dij)

’= fshareraw= Sj=1 

.Then fitness sharing is fi’= fi/mi 

.Fitness sharing with scaling is fi’= fi 

b / mi, b >= 1 

top-down view of a 2-d space 

raw 

fitness 

shared fitness 

without scaling 

side-view of the 2-d space 

with scaling using sufficiently large b 



A Dilemma 

.With low scaling factor: individuals won’t go to the real optimum 

because it’s not attractive 

.With high scaling factor: We may not be able to find all peaks, 

because a high scaling factor creates “super individuals”, even a 

very soft selection scheme won’t help 

. 

A possible solution: Anneal the factor b 

1. Start the evolution with a small b, e.g., b = 1, 

in order to explore and locate the peak regions 

2 Then increase b gradually to attract individuals to the optima 



Implicit Fitness Sharing (1) 

.The idea comes from an immune system: antibodies which best 

match an invading antigen receive the payoff for that antigen 

.Similar situation occurs in games: a strategy receives payoff 

when it achieves the best score against a test case 

.Implicit fitness sharing is most often used in learning. While 

(explicit) fitness sharing is done through individuals, implicit 

fitness sharing is test data based! 

.The algorithm for calculating fitness: 

For each data point i to be matched, do the following C times 

1. Select a sample of s 

individuals from the population 

2. Find the individual in the sample that achieves the highest 

score against the data point i 

3. This best individual receives the payoff. In the case of a tie, 

payoff is shared equally 



Implicit Fitness Sharing (2) 

.It has been shown that implicit and explicit fitness sharing have 

the same theoretical basis. s 

here plays the role of sshare in (explicit) 

fitness sharing 

.Larger C . 

better result but more time-consuming 

.Comparison between implicit and explicit sharing: they are better 

under different circumstances 

.Implicit fitness sharing covers optima more comprehensively, 

even when those optima have small basin of attraction, when 

the population is large enough for a species to form at each 

optimum 

.(Explicit) fitness sharing can find the optima with larger 

basins of attraction and ignore the peaks with narrow bases, 

when the population is not large enough to cover all optima 



Niching vs. Speciation 

.Although some people distinguish between the two, we will treat 

them as the same thing 

.If there is any difference: 

.niching is concerned more with locating peaks (basins of 

attraction), while speciation is more focused on actually 

converging to optima