Pareto Cooperative-Competitive Genetic Programming: A classification benchmarking study∗

A model for problem decomposition in Genetic Programming based classification is proposed consisting of four basic components: competitive coevolution, local Gaussian wrapper operators, evolutionary multiobjective (EMO) fitness evaluation, and an explicitly cooperative objective. The framework specifically emphasizes the relations between different components of the model. Thus, both the local wrapper operator and cooperative objective components work together to establish exemplar subsets against which performance is evaluated and the decomposition of the problem domain is achieved. Moreover, the cost of estimating fitness over multiple objectives is mitigated by the ability to associate specific subsets of exemplars with each classifier. The competitive coevolutionary model, in combination with a balanced sampling heuristic guides the identification of relevant exemplars and retention of the best classifiers; whereas the EMO component drives the application of selection and variation operators to maintain the diversity of the classifier population. Empirical evaluation is conducted over twelve data sets representative of different application domains, class distributions and feature counts. A benchmarking methodology is developed in order to provide comparison between the proposed model and the deterministic classifier paradigm in general, and the SVM method in particular. Moreover, by adopting performance metrics for classification, model complexity and training scalability, we are able to provide a more informative assessment of the relative merits of each model. ∗Extended version of paper published in Genetic Programming Theory and Practice VI: 2008, R. Riolo et al. (eds) Genetic and Evolutionary Computation Series – Copyright 2008 Springer-Verlag †Faculty of Computer Science, Dalhousie University. Halifax. NS. Canada.