Avaliação de topologia da rede neural e parâmetros do aprendizado de máquina por reforço para agentes jogadores de videogames

Abstract

Reinforcement machine learning has been an important area in the field of artificial intelligence due to its versatility and adaptability in complex settings, without the need of a specialist supervisor. The use of images as input signals makes the intelligent agent similar to how a human being makes decisions, using the image’s rich information. In reinforcement learning with neural networks, performance is highly dependant on the topology of the neural network and several lerarning parameters. Therefore, this aims to develop an intelligent agent for playingt Atari 2600 videogames, while also proposing changes to the topology and reinforcement learning parameters comapred to literature methods. Experimental results were successful in the evolution of the agent in maximizing rewards, which shows the proposed topology and parameters are viable.