Hibridização de Programação Genética Gramatical com estratégia evolutiva e evolução diferencial aplicada a problemas de regressão simbólica e de classificação

Abstract

Symbolic regression is the manipulation of mathematical expressions to find a function that best represents a dataset. Classification problems can be understood as the effect of arranging by classes a set of elements. Computational techniques were developed to solve these types of problems, Genetic Programming (GP) is one of them. One advantage of this technique is that it produces symbolic models that are easy to interpret. Grammar-based Genetic Programming (GGP) came up with the use of formal grammars to aid in the search for these models. One GGP problem is the coefficient adjustment as only va-lues from the grammar can appear within a model. Different hybridization forms of GGP are used in this work to solve this problem. Differential Evolution (DE) and Evolution Strategy (ES) are continuous optimization techniques, the objective here is to generate better solutions than a standard GGP, by the usage of these techniques to perform the adjustment of the coefficients. We tested 23 symbolic regression and 7 classification pro-blems to compare the performance of the proposed techniques with the standard GGP. Promising results compared to standard GGP were found.