@article{oai:ipsj.ixsq.nii.ac.jp:00107197, author = {中田, 雅也 and ピエール・ルカ・ランチ and 田島, 友祐 and 高玉, 圭樹 and Masaya, Nakata and Pier, LucaLanzi and Yusuke, Tajima and Keiki, Takadama}, issue = {2}, journal = {情報処理学会論文誌数理モデル化と応用(TOM)}, month = {Nov}, note = {本論文では,学習分類子システム(Learning Classifier System: LCS)において,学習する分類子数を削減するために,Compact Genetic Algorithmを用いた確率モデル型分類子生成法を提案する.提案分類子生成法は,1)分類子が持つ部分解の存在確率をモデル化することで不要な分類子の生成を抑制し,2) 従来の確率モデル型分類子生成法が適用困難であった強化学習問題クラスに適用可能である.教師あり学習問題(Multiplexer問題)と強化学習問題(Grid world問題)において,提案分類子生成法を導入したLCSを適用したところ,次の知見を得た.まず,1)提案LCSは従来LCSよりも,少ない学習回数で最適解を学習可能であり,2)従来LCSが学習した分類子数に対し,提案LCSは最小でも49%(最大で76%)削減した分類子数で学習可能であることを示した.したがって,提案分類子生成法は,最適解を持つ分類子を早期に生成可能であり,不要な分類子の生成を抑制可能であることを示した., This paper proposes a novel probability model based rule-discovery mechanism using Compact Genetic Algorithm for Learning Classifier System (LCS), which evolves classifiers based on an extracted attribute of classifier conditions, to reduce a size of classifiers are needed in LCS. The proposed rule-discovery mechanism can 1) generate good classifiers that conditions have good building blocks; and 2) solve both single-step problems and multi-step problems where conventional probability-model based rule discovery mechanisms are hard to be applied. This paper applies LCS with the proposed rule-discovery mechanism to both a single-step problem (the multiplexer problem) and a multi-step problem (the grid world problem). Experimental results show following implications: 1) the proposed LCS can reach optimal performances faster than a conventional LCS; and 2) it can reduce the size of classifiers by at least 49% of that of the conventional LCS. Our conclusion is that the proposed rule-discovery mechanism can generate optimal classifiers with fewer generations than the conventional rule-discovery mechanism, and that it can control generating inaccurate classifiers toward the rule reduction.}, pages = {1--16}, title = {Compact Genetic Algorithmを導入した学習分類子システムによる分類子数の削減}, volume = {7}, year = {2014} }