An Online Rule Set Compaction Strategy for the sUpervised Classifier System UCS

Abstract

Learning Classifier Systems (LCS) are an adaptive rule based class of algorithms driven by evolutionary mechanisms combined with machine learning. The goal of LCS is to create an entire population of rules, affording them the ability to learn iteratively and solve a given problem. The evolved population may contain many redundant or poor rules which can make interpretation and knowledge discovery by experts a considerable challenge. Therefore, it becomes essential to use rule compaction methods to achieve a balance between the maximum accuracy and a reduced population size. In an attempt to deal with the rule set reduction problematic, several rule compaction strategies were proposed as a post processing step to the algorithm. The aim of this work is to propose an Online Rule Compaction strategy for the sUpervised Classifier System (UCS). It is mainly based on clustering through the application of an adaptive version of the k-means algorithm.The proposed solution has showed promising results.