XCS with Bit Masks

Abstract

In this paper, a modified XCS is proposed to reduce the numbers of learned rules. XCS is a type of learning classifier systems and has been proven able to find accurate, maximal generalizations. However, XCS usually produces too many rules such that the readability of the classification model is greatly reduced. As a result, XCS users may not be able to obtain the desired knowledge or useful information from the learned rule set. In our attempt to handle this problem, a new mechanism, called bit masks, is devised in order to reduce the number of classification rules and therefore to improve the readability of the generated model. A series of n-bit multiplexer experiments, including 6-bit, 11-bit, and 20-bit multiplexers, to examine the performance of the proposed framework. For the problem composed of integertyped variables, two synthetic oblique datasets, Random-Data2 and Random-Data9, are adopted to compare the performance of XCS and that of the proposed method. According to the experimental results, XCS with bit masks can perform similarly as XCS on n-bit multiplexers and generates significantly fewer rules on integer-typed problems.