A regression-like classification system for geometric semantic genetic programming

Abstract

Geometric Semantic Genetic Programming (GSGP) is a recent variant of Genetic Programming, that is gaining popularity thanks to its ability to induce a unimodal error surface for any supervised learning problem. Nevertheless, so far GSGP has been applied to the real world basically only on regression problems. This paper represents an attempt to apply GSGP to real world classification problems. Taking inspiration from Per-ceptron neural networks, we represent class labels as numbers and we use an activation function to constraint the output of the solutions in a given range of possible values. In this way, the classification problem is turned into a regression one, and traditional GSGP can be used. In this work, we focus on binary classification; logistic constraining outputs in [0,1] is used as an activation function and the class labels are transformed into 0 and 1. The use of the logistic activation function helps to improve the generalization ability of the system. The presented results are encouraging: our regression-based classification system was able to obtain results that are better than, or comparable to, the ones of a set of competitor machine learning methods, on a rather rich set of real-life test problems.