Evaluating the Performance of the Deep Active Imitation Learning Algorithm in the Dynamic Environment of FIFA Player Agents

Abstract

Deep Learning is a state-of-the-art approach for machine learning using real-world or realist data. FIFA is a soccer simulation game that provides a very realistic environment, but which has been relatively poorly explored in the context of learned game-playing agents. This paper explores the Deep Active Imitation (DAI) learning strategy applied to a dynamic environment in FIFA game. DAI is a segment of Imitation Learning, which consists of a supervised Deep Learning training strategy where the agents learn by observing and replicating human experts' behavior. Noteworthy here is that such learning strategy has only been validated in static navigation scenarios in the sense that the environment is changed only through the actions of the agent. In this way, the main objective of the present work is to investigate the efficacy of DAI to cope with a dynamic FIFA scenario named confrontation mode. The agents were evaluated in terms of in-game score through tournaments against FIFA's engine. The results show that DAI performs well in the confrontation mode. Thus, this work indicates that such learning strategy can be used to solve complex problems.