Otimização de hiperparâmetros em algoritmos de arvore de decisão utilizando computação evolutiva

Abstract

Some algorithms in machine learning are parameterizable, they allow the configuration of parameters in order to increase the performance in some tasks. In most cases, these parameters are empirically found by the developer. Another approach is to use some optimization technique to find an optimized set of parameters. The aim of this project is the application of evolutionary algorithms, Genetic Algorithm (GA), Fluid Genetic Algorithm (FGA) and Genetic Algorithm using Theory of Chaos (GATC) to optimize the search for hyperparameters in decision tree algorithms. This work presents some satisfactory results within the data set tested, where the Classification and Regression Trees (CART) algorithm was used as a classifier algorithm for the tests. In these, the decision trees generated from the default values of the hyperparameters are compared with those optimized by the proposed approach. We has tried to optimize the accuracy and final size of the generated tree, which were successfully optimized by the proposed algorithms.