Re-Labeling Differential Evolution for Combinatorial Optimization and Interactive Evolutionary Computation

Ryohei Funaki, Hirotaka Takano, Junichi Murata

Research output: Contribution to journalArticlepeer-review

Abstract

The authors proposed a new differential evolution technique, Re-Labeling Differential Evolution (RLDE), which, in this paper, is refined and evaluated in the context of interactive solution of combinatorial optimization problems. Many of the practical design problems such as web page layout design and room lighting design are combinatorial optimization problems where the numerical evaluations are not available. The evaluation should be made by humans. There are two essential properties necessary for the solution methods to the above problems: (1) interaction between the methods and the users to extract human evaluations accurately without imposing too much burden on the users, and (2) abilities to solve combinatorial optimization problems. Interactive differential evolution (IDE) techniques possess the first property because they utilize pairwise comparisons but lack the second property, while interactive genetic algorithms have the second property but not the first one. RLDE is an extended algorithm of IDE so that it can solve combinatorial optimization problems. In differential evolution (DE), solution candidates are represented by numerical values. In combinatorial optimization problems, however, the numerical values are only labels to distinguish the components to be combined, and there are no structural relationships such as large/small and far/near among them which the DE relies on. RLDE collects information on the problem while it searches for the solution, and, based on the obtained information, re-labels the components so that the DE algorithm works more efficiently. RLDE was originally proposed as a technique for simple combinatorial optimization. In this paper, the authors extend RLDE to permutation-based combinatorial optimization. The performance of RLDE in terms of the burden on the users and the quality of the obtained solutions is evaluated and compared with the other techniques in numerical experiments.
Original languageEnglish
Pages (from-to)18-25
Number of pages8
JournalSICE Journal of Control, Measurement, and System Integration
Volume9
Issue number1
DOIs
Publication statusPublished - 2016

Fingerprint

Dive into the research topics of 'Re-Labeling Differential Evolution for Combinatorial Optimization and Interactive Evolutionary Computation'. Together they form a unique fingerprint.

Cite this