VARIABLE NEIGHBORHOOD PROGRAMMING FOR EVOLVING DISCRIMINENT FUNCTIONS WITH DYNAMIC THRESHOLDS

摘要

This paper aims to develop a new automatic programming system to solve multi-class classification problem. Our method is based on a recently proposed algorithm called variable neighborhood programming (VNP). VNP is a solution-based metaheuristic which evolves programs and classifiers. Numeric expression is one of classifier representation in automatic programming algorithms. In the case of binary classification, a translation can simply performed according to the sign of the output (negative: class 1, positive: class 2). However, when the number of classes is higher, setting the appropriate boundary values and fixing the optimal class order is more complicated. To solve this problem, we introduce a new local search which aims to find dynamically the more adequate thresholds for each classifier. To include the developed local search within VNP (and also within any automatic programming algorithm), a multidimensional schema is proposed to ensure the simultaneous optimization of the discriminant function (tree), the class order, and the corresponding boundaries. Experimental results on three real data sets show that the proposed approach is a good automatic programming tool, able to build a full classifier with high accuracy in a single run.