Aprendizagem supervisionada na predição de curvas de valores fitness em algoritmos genéticos

Abstract

This paper continues the efforts of a recent approach that proposed the prediction of fitness value curves in the evolutionary process of Genetic Algorithms (GAs). The current work proposes to deepen and suggest new methods and environments in order to understand, among other aspects, the influences of the size and quality of the dataset, as well as the characteristics of the problems used. It was found that the new methods contributed to a more accurate understanding, allowing to highlight some difficulties that the model is susceptible to, mainly caused by little executions data available for training when the problem solved by the GA does have many local minimums.