Representational effects in a simple classifier system
Sandip Sen · 1994
Classifier gystems are rule-based machine learning systems in which natural genetics motivate the learning mechanism.They have been applied to a number of supervised classification problems in recent years with very competitive results compared to other supervised learning algorithms.The effectiveness of learning in such systems largely depends on the representation used for the training and testing instances, as well as that used for the classification rules.In this paper, we present the effects of alternative representation of classification rules in NEWBOOLE, a simple stimulusresponse classifier system, when applied to a well-known supervised concept learning problem.We present an analysis of the concept learning problem as faced by a pattern matching rule-based system like NEWBOOLE.A representation of rules influenced by domain knowledge is shown to improve the performance of NEW-B()OLE over a straightforward representation both in terms of the accuracy of prediction and the speed of learning.(;lassitication accuracy obtained on the t.esting data sets is found to be better than that reported in contemporary literature.