previous  home  search  LaTeX (120kb)   PostScript (269kb)   PDF (164kb)   contact    up    next  

Fitness Uniform Selection to Preserve Genetic Diversity


Author: Marcus Hutter (2001-2002)
Comments: 13 pages, 1 figure
Subj-class: Artificial Intelligence; Learning; Distributed Computing
ACM-class: I.2; I.2.6; I.2.8; F.2
Reference: Proceedings of the 2002 Congress on Evolutionary Computation (CEC-2002) 783--788, IEEE
Report-no: IDSIA-01-01 and cs.AI/0103015
Paper: LaTeX (120kb)  -  PostScript (269kb)  -  PDF (164kb) 
Slides: PostScript - PDF
C++ Code:FussDD.cpp - FussDD.h - FussTSP.cpp - FussTSP.h
Review/Survey: in the Technology Reseach News Magazine (cached)
Extended Abstract at ECCO-XV conference

Keywords: Evolutionary algorithms, fitness uniform selection strategy, preserve diversity, local optima, evolution, correlated recombination, crossover.

Abstract: In evolutionary algorithms, the fitness of a population increases with time by mutating and recombining individuals and by a biased selection of more fit individuals. The right selection pressure is critical in ensuring sufficient optimization progress on the one hand and in preserving genetic diversity to be able to escape from local optima on the other. We propose a new selection scheme, which is uniform in the fitness values. It generates selection pressure towards sparsely populated fitness regions, not necessarily towards higher fitness, as is the case for all other selection schemes. We show that the new selection scheme can be much more effective than standard selection schemes.


 previous  home  search  LaTeX (120kb)   PostScript (269kb)   PDF (164kb)   contact    up    next  

Table of Contents

  1. Introduction
    • Evolutionary algorithms (EA)
    • Standard selection schemes (STD)
    • The problem of the right selection pressure
    • The main idea
    • Contents
  2. Fitness Uniform Selection Strategy (FUSS)
    • The problem of local optima
    • The fitness uniform selection scheme (FUSS)
    • Asymptotically fitness uniform population
    • Fitness gaps and continuous fitness
    • Mutation and recombination
  3. Properties of Fuss
    • FUSS effectively favors fit individuals
    • No takeover in FUSS
    • FUSS may also favor unfit individuals
    • Distribution within a fitness level
    • Steady creation of individuals from every fitness level
    • Transformation properties of FUSS
  4. A Simple Example
    • Simple 2D example
    • Random search
    • Random search with crossover
    • Mutation
    • Standard selection with crossover
    • FUSS
    • FUSS with crossover
    • Simple 3D example
    • Numerical results
  5. Recombination
    • Worst case analysis without recombination
    • Quadratic slowdown due to recombination
    • Scale independent pair selection
    • Properties of $p(f,f')$
  6. Continuous Fitness Functions
    • Effective discretization scale
    • FUSS
    • Discussion
  7. Summary & Conclusions
 previous  home  search  LaTeX (120kb)   PostScript (269kb)   PDF (164kb)   contact    up    next  

BibTeX Entry

@Article{Hutter:01fuss,
  author =       "Marcus Hutter",
  number =       "IDSIA-01-01",
  institution =  "Istituto Dalle Molle di Studi sull'Intelligenza Artificiale",
  title =        "Fitness Uniform Selection to Preserve Genetic Diversity",
  month =        jan,
  year =         "2001",
  pages =        "13 pages",
  address =      "Manno(Lugano), CH",
  journal =      "Submitted to IEEE Transactions Evolutionary Computation",
  keywords =     "Evolutionary algorithms, fitness uniform selection strategy,
                  preserve diversity, local optima, evolution,
                  correlated recombination, crossover.",
  url =          "http://www.hutter1.net/ai/pfuss.htm",
  url2 =         "http://arxiv.org/abs/cs.AI/0103015",
  ftp =          "ftp://ftp.idsia.ch/pub/techrep/IDSIA-01-01.ps.gz",
  abstract =     "In evolutionary algorithms, the fitness of a population increases
                  with time by mutating and recombining individuals and by a biased
                  selection of more fit individuals. The right selection pressure is
                  critical in ensuring sufficient optimization progress on the one
                  hand and in preserving genetic diversity to be able to escape from
                  local optima on the other. We propose a new selection scheme,
                  which is uniform in the fitness values. It generates selection
                  pressure towards sparsely populated fitness regions, not
                  necessarily towards higher fitness, as is the case for all other
                  selection schemes. We show that the new selection scheme can be
                  much more effective than standard selection schemes.",
}
      
 previous  home  search  LaTeX (120kb)   PostScript (269kb)   PDF (164kb)   contact    up    next