Aplicação de um modelo Neuroevolutivo para simular jogadores : um estudo de caso com o jogo Ultimate Guitar Show

Abstract

This work presents a machine learning model that, after specialized training, obtains the necessary skills to play the Ultimate Guitar Show game (developed by the author). For this, neural networks were used combined with genetic algorithms in a neuroevolutionary approach. In the context of this work, genetic algorithms are used to define synaptic weights during the training of a deep neural network. The neural network, in turn, returns a binary value that represents a decision in the game. Each chromosome is tested individually and, through a specific calculation, its fitness value (level of aptitude in playing) is defined and attributed to it. Seven rounds of training were carried out until the best genetic configuration was reached in the last one, resulting in the production of several extremely specialized chromosomes in the game. After all the training phases, the chromosome generated with the best fitness was then tested in a neural network within the real game environment. Such a test, carried out both with the phase used in training and other phases that the neural network did not know about, proved the occurrence of learning, since in all these different phases the neural network presented the same level of accuracy in playing autonomously