Learning Classifier Systems BRANDEN PAPALIA, MICHAEL STEWART, JAMES PATRICK FACULTY OF ENGINEERING, COMPUTING AND MATHEMATICS
The University of Western Australia Introduction Learning Classifier Systems (LCS) are rule-based online machine learning systems LCS combine machine learning techniques with evolutionary learning techniques Generally consists of a set of rules (population of classifiers), a rule evaluation mechanism, and a rule evolution mechanism Several types, namely: Pittsburgh-style Michigan-style –XCS –ZCS
The University of Western Australia History Introduced by John H Holland in 1975 First application in the 1980s Further developed by Stewart W Wilson in Accuracy-based XCS system Strength-based ZCS system Pittsburgh LCS introduced by Kenneth Dejong
The University of Western Australia Interaction of Agent/Environment Learning Classifier System Environment Reward State Action Browne and Urbanowicz, 2013
The University of Western Australia Rules Based on traditional computer logic IF condition THEN action Rules depend on problem environment Use fitness (ZCS) or accuracy (XCS) to evaluate rules Break problem into sub problems
The University of Western Australia Rule Structure Condition Generally looks for certain kinds of messages Action Specifies a message to be sent when rule is satisfied Reward Used in reinforcement learning Statistics
The University of Western Australia Rule evaluation mechanism
The University of Western Australia Michigan vs. Pittsburgh Pittsburgh is rarely used – Michigan is the prominent LCS type Michigan ZCS (strength based) XCS (accuracy based) Pittsburgh evolves one global problem solution, searches for small sets of rules Michigan is distributed, searches for larger distributed sets of rules
The University of Western Australia LCS as a Map Generator Browne and Urbanowicz, 2013 Browne and Urbanowics 2013
The University of Western Australia General algorithm for LCS operation Browne and Urbanowics, 2013 Browne and Urbanowics 2013
The University of Western Australia With Example Binary Encodings Browne and Urbanowicz, 2013 Browne and Urbanowics 2013
The University of Western Australia Learning Browne and Urbanowicz, 2013 Browne and Urbanowics 2013
The University of Western Australia A Simple Example Takadama Labs 2015
The University of Western Australia Applicable Problem Types Most Prediction type problems Classification type Reinforcement learning Function Approximation Booker, Colombetti, Dorigo, et, al. 2000
The University of Western Australia Problem Characteristics Where LCS Performs Well Multimodal Lack of separation High dimensionality Epistatisistic Characteristics
The University of Western Australia Contemporary Research into LCS' Scalability Anticipation Future Research Complex Rule Dependencies Butz, 2007
The University of Western Australia References "Learning Classifier Systems: Introducing the User-friendly Textbook", Browne and Urbanowicz 2013 "Design and Analysis of Learning Classifier Systems", Drugowitsch, Springer 2006 "Combining Gradient-Based With Evolutionary Online Learning: An Introduction to Learning Classifier Systems, Butz 2007 "Learning Classifier System with Adaptive Action Mapping " Takadama Lab 2015 “Learning Classifier Systems: Introducing the User-friendly Textbook” Browne and Urbanowicz, 2013 “Design and Analysis of Learning Classifier Systems, Drugowitsch, Springer 2006 “Combining Gradient-Based With Evolutionary Online Learning: An Introduction to Learning Classifier Systems”, Butz 2007