Evolving Multiple Discretizations with Adaptive Intervals for a Pittsburgh Rule-Based Learning Classifier System

摘要

One of the ways to solve classification problems with real-value attributes using a Learning Classifier System is the use of a discretization algorithm, which enables traditional discrete knowledge representations to solve these problems. A good discretization should balance losing the minimum of information and having a reasonable number of cut points. Choosing a single discretization that achieves this balance across several domains is not easy. This paper proposes a knowledge representation that uses several discretization (both uniform and non-uniform ones) at the same time, choosing the correct method for each problem and attribute through the iterations. Also, the intervals proposed by each discretization can split and merge among them along the evolutionary process, reducing the search space where possible and expanding it where necessary. The knowledge representation is tested across several domains which represent a broad range of possibilities.