Skip to Main content Skip to Navigation

Etude et Exploitation des Réseaux de Neutralité dans les Paysages Adaptatifs pour l'Optimisation Difficile

Abstract : The concept of fitness landscape (or adaptive landscape) was introduce par S. Wright in the field of evolutionary biology in 1930's. It is one of the most relevant to explain the evolution of individuals. In the field of combinatorial optimization by metaheuristic, it is also used and allows to study the link between
geometrical description of optimization problem and the dynamic of search algorithms. Two geometries of landscape which correspond to two dynamics of search have been studied. The multimodal geometry of landscape is related to the presence of local optima, where the search dynamic is a succession of adaptive walk toward better solutions and degradation of performance.
The geometry of neutral fitness landscape, point out in molecular evolution by neutral theory of Motoo Kimura, is related to presence of plateaus ; the dynamic of search is characterized by random drift interrupted by the discover of rare better solution. This thesis propose to deeper study neutral fitness landscapes in the context of optimization and to design new metaheuristics according to those landscapes.

This thesis is composed by four parts. In the first one, we present the main results about fitness landscapes and more particularly about neutral fitness landscapes. In the second part, we develop the concept of neutral set by introducing the notion of 'fitness cloud' which allows to study the correlation of performance between two neighbor solutions and we measure this correlation on 'embedded fitness landscapes' as an extension of NK landscapes and Max-SAT problems. In the third part, we summarize the set of measures on neutral networks and we propose the new measure.
Experimental study is performed on three family of landscapes for which the neutrality is and two classical problems. Then, a new metaheuristic adapted of neutral fitness landscapes inspired by the new measure is proposed and evaluated on different landscapes. We studied the massively neutral fitness landscapes from the learning problem of a rule of cellular automata which perform the density task, in order to improve the best metaheuristics known.
Complete list of metadata
Contributor : Sébastien Verel <>
Submitted on : Tuesday, July 3, 2007 - 8:49:06 PM
Last modification on : Wednesday, April 14, 2021 - 4:06:01 PM
Long-term archiving on: : Thursday, April 8, 2010 - 7:46:25 PM


  • HAL Id : tel-00159727, version 1



Sébastien Verel. Etude et Exploitation des Réseaux de Neutralité dans les Paysages Adaptatifs pour l'Optimisation Difficile. Autre [cs.OH]. Université Nice Sophia Antipolis, 2005. Français. ⟨tel-00159727⟩



Record views


Files downloads