《《lcs研究概述》ppt课件》由会员分享，可在线阅读，更多相关《《lcs研究概述》ppt课件（47页珍藏版）》请在金锄头文库上搜索。
1、Research on Learning Classifier System,学习分类系统研究概述,Outline,Introduction Definition，History Basic Idea of LCS Types，Approaches Our Current Progress What we have done? Hot Issues and Future Direction What is a promising future research direction?,Outline,Introduction Definition，History Basic Idea of LC
2、S Types，Approaches Our Current Progress What we have done? Hot Issues and Future Direction What is a promising future research direction?,Adaptive rule-based production system Set of rules Trial and error: 强化学习 通过环境的反馈来调整自身的行为 Survival of the fittest: 遗传算法 以遗传算法来探索、发现规则,Introduction to LCS,Rule A,Ru
3、le B,Reinforcement Learning,Environment,RL Agent,action,Reward & State,Genetic Algorithm,Population,(2),Population,Population,New Individual,Good Individual,(1),Individual,Selection,Reproduction,Components in LCS,Components in LCS,Components in LCS,Brief View of LCS History,1971 Holland首次提出分类系统概念,19
4、78 Holland正式确立学习分类系统名称，并提出大概框架,1988 Holland定义标准框架(太复杂) LCS研究停滞,1995 Wilson进一步提出XCS，从此LCS的研究进入新的阶段,1994 Wilson简化了标准LCS，提出更易实现的ZCS,1998 Stolzmann提出不同于传统LCS的A-LCS，新的方向,Relative,Conferences and Magazines Genetic and Evolutionary Computation Conference (GECCO) International Workshop on Learning Classifie
5、r Systems (IWLCS） SEAL, Evolutionary Computation/IEEE Transaction on EC Papers and Applications H. Ishibuchi. Fuzzy Genetics-Based Machine LearningSEAL 2012 Pier Luca Lanzi. XCS with Adaptive Action Mapping SEAL 2012 R. Urbanowicz. Instance-Linked Attribute Tracking and Feedback for Michigan-Style S
6、upervised Learning Classier SystemsGECCO2012 M. Iqbal. Extracting and Using Building Blocks of Knowledge in Learning Classier SystemsGECCO2012 ,Outline,Introduction Definition，History Basic Idea of LCS Types，Approaches Our Current Progress What we have done? Hot Issues and Future Direction What is a
7、 promising future research direction?,Brief View of LCS History,1971 Holland首次提出分类系统概念,1978 Holland正式确立学习分类系统名称，并提出大概框架,1988 Holland定义标准框架(太复杂) LCS研究停滞,1995 Wilson进一步提出XCS，从此LCS的研究进入新的阶段,1994 Wilson简化了标准LCS，提出更易实现的ZCS,1998 Stolzmann提出不同于传统LCS的A-LCS，新的方向,Hollands LCS,缺陷： 1.无节制使用遗传算法 2.桶队列算法的依赖性,规则 条件
8、/动作/预测 匹配集M 动作选择 动作集A,Brief View of LCS History,1971 Holland首次提出分类系统概念,1978 Holland正式确立学习分类系统名称，并提出大概框架,1988 Holland定义标准框架(太复杂) LCS研究停滞,1995 Wilson进一步提出XCS，从此LCS的研究进入新的阶段,1994 Wilson简化了标准LCS，提出更易实现的ZCS,1998 Stolzmann提出不同于传统LCS的A-LCS，新的方向,Wilsons XCS,最重要的改进部分： 重新定义了适应度计算方法 Hollands LCS： 规则的权值 Wilsons
9、 XCS： 引入了新的参数通过计算精确度来度量遗传算法,Brief View of LCS History,1971 Holland首次提出分类系统概念,1978 Holland正式确立学习分类系统名称，并提出大概框架,1988 Holland定义标准框架(太复杂) LCS研究停滞,1995 Wilson进一步提出XCS，从此LCS的研究进入新的阶段,1994 Wilson简化了标准LCS，提出更易实现的ZCS,1998 Stolzmann提出不同于传统LCS的A-LCS，新的方向,Stolzmanns ACS,Model-Free LCSs ZCS/XCS No knowledge abou
10、t result of actions Model-Based LCS Anticipatory classifier systems(ACS,1998) Anticipatory learning classifier systems(ACS2,2000) Knowledge about result of actions,Two Approaches,(1) Michigan Approach: Search for good rules (2) Pittsburgh Approach: Search for a good rule combination,Champions=Good p
11、layers +Good cooperation,Michigan Approach,Fitness Evaluation of Each Rule Direct Optimization of Rules New rules are generated from good rules Indirect Search for a Good Rule Set A set of good rules is not necessarily a good rule set,Rule A,Rule C,Rule E,Rule G,Rule B,Rule D,Rule F,Rule H,Pittsburg
12、h Approach,Fitness Evaluation of Each Sub Rule Set Direct Optimization of Rule Sets New rule sets are generated from good rule sets Indirect Search for Good Rules Good rules in a poor rule set cannot survive,Rule A Rule B Rule C,Rule D Rule E Rule F,Rule G Rule H Rule I,Rule J Rule K Rule L,Michigan
13、-Pittsburgh Hybrid Approach,H. Ishibuchi et al. Hybridization of Fuzzy GBML Approaches for Pattern Classification Problems, IEEE T-SMC Part B (2005),Outline,Introduction Definition，History Basic Idea of LCS Types，Approaches Our Current Progress What we have done? Hot Issues and Future Direction What
14、 is a promising future research direction?,Improvement of LCS,Sub-LCS,LCSE,Rule A,Rule C,Rule E,Rule B,Rule B,Rule D,Rule F,Rule A,ability,readability,Sub-LCS,Sub-LCS,Ensemble Method,Parallel ensemble Bagging，Random subspace，Random forest create diverse base learners by introducing randomness Sequen
15、tial ensemble Adaboost create base learners by complementarity,LCSE: LCS Ensemble(Bagging),LCSE: LCS Ensemble(Boosting),Compact Rule Set,(Suppose simplest conditions) 2-D Problem: 32 = 9 rules 4-D Problem: 34 = 81 rules 6-D Problem: 36 = 729 rules 8-D Problem: 38 = 6,561 rules 10-D Problem: 310 = 59
16、,049 rules ,lack of readability traditional CRA is too complicated,Compact Rule Set,Yang Gao, Lei Wu, Joshua Zhexue Huang. Ensemble Learning Classifier System and Compact Ruleset. In: Proceedings of the 6th International Conference on Simulated Evolution and Learning. LNCS4247, pp:42-49, 2006.,Compact Rule Set,Yang Gao, Lei Wu, Joshua Zhexue Huang. Ensemble Learning Classifier System and Compact Ruleset. In: Proceedings of the 6th Intern
