Abstract

Abstract. In designing non-linear classifiers, there are important trade-offs to be made between predictive accuracy and model comprehensibility or complexity. We introduce the use of Genetic Programming to generate logistic polynomial models, a relatively comprehensible non-linear parametric model; describe an efficient twostage algorithm consisting of GP structure design and Quasi-Newton coefficient setting; demonstrate that Niched Pareto Multiobjective
Genetic Programming can be used to discover a range of classifiers with different complexity versus “performance” trade-offs; introduce a technique to integrate a new “ROC (Receiver Operating Characteristic) dominance” concept into the multiobjective setting; and suggest some modifications to the Niched Pareto GA for use in Genetic Programming. The technique successfully generates classifiers with diverse complexity and performance characteristics.