Problem Driven Machine Learning by Co-evolving Genetic Programming Trees and Rules in a Learning Classifier System

Abstract

A persistent challenge in data mining involves matching an applicable as well as effective machine learner to a target problem. One approach to facilitate this process is to develop algorithms that avoid modeling assumptions and seek to adapt to the problem at hand. Learning classifier systems (LCSs) have proven themselves to be a flexible, interpretable, and powerful approach to classification problems. They are particularly advantageous with respect to multivariate, complex, or heterogeneous patterns of association. While LCSs have been successfully adapted to handle continuous-valued endpoint (i.e. regression) problems, there are still some key performance deficits with respect to model prediction accuracy and simplicity when compared to other machine learners. In the present study we propose a strategy towards improving LCS performance on supervised learning continuous-valued endpoint problems. Specifically, we hypothesize that if an LCS population includes and co-evolves two disparate representations (i.e. LCS rules, and genetic programming trees) than the system can adapt the appropriate representation to best capture meaningful patterns of association, regardless of the complexity of that association, or the nature of the endpoint (i.e. discrete vs. continuous). To successfully integrate these modeling representations, we rely on multi-objective fitness (i.e. accuracy, and instance coverage) and an information exchange mechanism between the two representation ‘species’. This paper lays out the reasoning for this approach, introduces the proposed methodology, and presents basic preliminary results supporting the potential of this approach as an area for further evaluation and development.