Experimentos com Aprendizado por Reforço em Cenário de Combate de StarCraft

Abstract

In order to improve the artificial intelligence (AI) of real-time strategy (RTS) games, the AI international community has encouraged researches the context of several problems on the RTS games environment. This is due to the fact the RTS games offer to researchers an excellent labs due to the complexity and dynamicity of the simulation of realistic, controlled and observable battle fields. Those environment, such as the game StarCraft, witch is a benchmark of the AI international community, require complex decision make solutions. Witch involves problems like resource management, uncertainty reasoning, collaboration and unities coordination, opponent learning and modelling, etc, in order to reach their goal in the best way. For those reasons, this work propose to analyze and improve the application of reinforcement learning techniques in a simplified StarCraft combat scenario. In order to do this, this work is based in a work published on IEEE Conference on Computational Intelligence and Games, through the experimentation of various reinforcement learning parameters as well as modification of the original environment representation model used the published work. The results obtained from the experiments show a significant improvement on the learning efficiency and quality. This ends up reflecting in a augmentation of the victories average on the simulated environment.