Skip to Main content Skip to Navigation
Theses

Stratégies d'exploration de paysages de fitness : application à la résolution approchée de problèmes d'optimisation combinatoire

Abstract : Many combinatorial optimization problems are hard to solve and in many cases, exact approaches are impracticable. Among partial search algorithms, metaheuristics are generic algorithms, widely studied in the literature. Their ability to find good solutions varies in function of the problems’ nature et data composing problem instances, and studying efficiently the dynamics of such algorithms is challenging, especially for large instances. We restrain our metaheuristic study to local search algorithms. Basic mechanisms are studied to improve their understanding and assess their ability to find good solutions. We abstract optimization problems into fitness landscapes, thanks to a neighborhood relation between solutions, in order to analyze the dynamics of methods in function of several landscapes characteristics.We study the navigation on these landscapes, firstly by constraining moves to be strictly improving. In particular, we propose the expansion criterion to guide the search process and assess its relevance to guide climbers through good solutions. Variants approximating this principle are proposed and studied, leading to many trade-offs between the ability to find good solutions and the computational cost making them integrable into more complex metaheuristics. Last, we study partial neighborhood local searches, which accept deteriorating moves. In this context, experiments show that simple pivoting rules are sufficient to attain good trade-offs between intensification and diversification and thus reaching good solutions.
Document type :
Theses
Complete list of metadata

Cited literature [207 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-02469501
Contributor : Abes Star :  Contact
Submitted on : Thursday, February 6, 2020 - 3:41:23 PM
Last modification on : Friday, June 19, 2020 - 1:18:05 PM
Long-term archiving on: : Thursday, May 7, 2020 - 3:58:37 PM

File

these_TARI.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-02469501, version 1

Citation

Sara Tari. Stratégies d'exploration de paysages de fitness : application à la résolution approchée de problèmes d'optimisation combinatoire. Informatique et langage [cs.CL]. Université d'Angers, 2019. Français. ⟨NNT : 2019ANGE0013⟩. ⟨tel-02469501⟩

Share

Metrics

Record views

258

Files downloads

272