A209 種分類を用いた進化計算によるセルオートマトン・ルール自動設計(最適設計関連)

概要

論文の詳細を見る
Cellular Automata (CA) are one of the visualization and analyzation tool of complex phenomena. When we reproduce the phenomena with CA,the equation of motion for a CA (CA-Rules) shall be designed according to behavior pattern of each cells. In the case of difficult situation in the CA-Rules designing (such as behavior pattern is more intricacy, analysis of the behavior pattern is impossible), Automatic Designing CA-Rules Technic is the essential skills. Density Classification is the well-studied task known as designing problem for one dimensional CA. The task is to determine whether p_θ, the fraction of 1s in teh initial configuration (IC), is greater than or less than a critical value ρ_c. If ρ_o > ρ_c, the entire lattice should relax to a fixed point of all 1s ; otherwise it should relax to a fixed point of all Os. Designing CA-Rules in Density Classification is difficult extremely, because the space of CA-Rules that are possible is too large and the accurate rule that set off a global behavior won't be fixed. Gacs-Kurdyunov-Levin-rule (GKL-rule) is the hand-designed rule that can fix the Density Classification task at a higher performance. The GKL-rule has operated 82.7%'s accurately in 10000's ICs sampled at random. In [Mitchell 94], M. Mitchell reported results of Designing CA-Rules used Simple Genetic Algirithm (SGA) to perform Density Classification. Desigined CA-Rules (SGA-rule) in [Mitchell 94] has operated 76.1 %'s accurately in 10000's ICs sampled at random, and this rule can't overreach the GKL-rule. We describe an alternative approach of Evolutionary Computations (ECs) to the design of CA that can perform computations requiring global coordination. On recent works, Coevolutionary Learning was used to evolve CA-Rules for the density classification. In this cases, the EC discovered rules that didn't give rise to high performance and sophisticated strategy. Some reason presented this experiment seem to prevent continuous progress in evolutionary serch. The major reason is that the dynamics of the search performed by the two coevolving populations doesn't drive individuals to the domain of the state space that contains most promising solutions because there is no "high-level" strategy to play that role. Our approach proposes a coevolutionary framework in which those two issues are addressed as follows : (1) Classification of CA-Rules into some of "species" by the use of its state of affairs in CA task. (2) Allocating the perfect "Fitness-Function" for each species so as to aggregate own strategy each other. We analyze the results of our methods on computer experiments, and weigh against existing techniques in reference to general purpose.
社団法人日本機械学会の論文
2001-11-14

A209 種分類を用いた進化計算によるセルオートマトン・ルール自動設計(最適設計関連)

スポンサーリンク

概要

著者

関連論文

スポンサーリンク