Applying the XCS Learning Classifier System to continuous-valued data-mining problems

This thesis represents an in depth investigation into the issues raised by the iterative nature of the data-mining process and, in particular, the use of the XCS Learning Classifier System with continuousvalued data-mining problems. The XCS Learning Classifier System has been shown to have the capability for data-mining through rule induction, that is, a technique by which various characteristics of a given problem space may be deduced and presented to the user in a readable format. There should exist a close interaction between user and machine-based adaptive search process whereby information provided by the search process helps the user to develop a better understanding of the problem domain. In addition, it is hoped that the user’s knowledge and intuition can be captured through an iterative refinement of the problem space. Three continuous-valued test environments are used to demonstrate how effective the XCS Learning Classifier System can learn to describe these environments. In addition, the issue of presenting human-readable information about these problem spaces to the user is addressed as well as an investigation into any capabilities the system has to deal with changes in the underlying environment that may have been made by a user. In particular, the thesis introduces two novel approaches to learning motivated by the need to learn fairly accurate models quickly. The first approach is a simplification of the classifier system’s update mechanism which proves to be advantageous in certain types of environment. The second is the inclusion of memetic learning within the update mechanism which also proves to be beneficial in terms of increased learning-speed.