Improved runtime bounds for the univariate marginal distribution algorithm via anti-concentration

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Colleges, School and Institutes

Abstract

Unlike traditional evolutionary algorithms which produce offspring via genetic operators, Estimation of Distribution Algorithms (EDAs) sample solutions from probabilistic models which are learned from selected individuals. It is hoped that EDAs may improve optimisation performance on epistatic fitness landscapes by learning variable interactions. However, hardly any rigorous results are available to support claims about the performance of EDAs, even for fitness functions without epistasis. The expected runtime of the Univariate Marginal Distribution Algorithm (UMDA) on OneMax was recently shown to be in O (nX log X) [9]. Later, Krejca and Witt proved the lower bound Q (Xs/n + n log n) via an involved drift analysis [16]. We prove a O (nX) bound, given some restrictions on the population size. This implies the tight bound 8 (n log n) when X = O (log n), matching the runtime of classical EAs. Our analysis uses the level-based theorem and anti-concentration properties of the Poisson-binomial distribution. We expect that these generic methods will facilitate further analysis of EDAs.

Details

Original languageEnglish
Title of host publicationGECCO '17 Proceedings of the Genetic and Evolutionary Computation Conference
Publication statusPublished - 1 Jul 2017
EventGECCO 2017:: The Genetic and Evolutionary Computation Conference - Berlin, Germany
Duration: 15 Jul 201719 Jul 2017
http://gecco-2017.sigevo.org/index.html/HomePage

Conference

ConferenceGECCO 2017:
CountryGermany
CityBerlin
Period15/07/1719/07/17
Internet address

Keywords

  • Runtime Analysis, Level-based Analysis, Estimation of Distribution Algorithms