Análise comparativa de técnicas de neuroevolução no aprendizado do jogo Chrome Dino.

Abstract

Neuroevolution (NE), a field of Computer Science that uses Evolutionary Computation (EC) techniques to optimize and train Artificial Neural Networks (ANNs), presents itself as a robust approach for creating autonomous agents for games. Defining the genetic operators used is an important step in developing a NE method, due to their influence on the ability to avoid premature convergence of the ANN model. In this context, this work presents the results of a comparative performance analysis between the canonical Genetic Algorithm (GA) with elitism and Rechenberg’s Evolution Strategy (ES) (μ, λ), applied to the training of a neural network for learning the game Chrome Dino. To this end, the Dinossauro-Google codebase was extended, creating classes that implement both the compared EC methods and the Strategy design pattern, in order to allow for flexible switching of genetic operators during the tests performed. The results of the final experiments indicated that the GA approach achieved the highest fitness peaks and the highest final mean, while the ES approach showed more consistent performance among the neuroevolutions tested. A two-tailed Mann-Whitney U test, with a significance level of α = 0.05, however, found insufficient evidence to claim that the final fitness distributions of the two approaches are statistically different. As contributions, this work offers quantitative evidence regarding the impact of the choice of evolutionary algorithm and the mutation hyperparameter on the final fitness distributions of neuroevolution applied to the Chrome Dino game, in addition to providing a reusable extension of the Dinossauro-Google codebase, where the use of the Strategy design pattern facilitates the investigation of new genetic operators in future research.