Deep Active Imitation Learning in FIFA Free-Kicks Player Platforms Based on Raw Image and Object Detection State Representations

摘要

This paper investigates the use of the Deep Active Imitation (DAI) learning strategy in single and multi-agents player platforms for FIFA game. DAI is a supervised Deep Learning training strategy in which previously trained agents evolve by mimicking human experts' behavior through active learning. Such learning strategy has only been validated in single-agent scenarios where the environment is perceived through raw images. FIFA, in spite of being a successful soccer simulation video game representing a very realistic and challenging case study, has been relatively poorly explored in the context of Machine Learning research. Then, the main objective of the present work is to investigate the efficacy of DAI to cope with single and multi-agents FIFA platforms in which the environment is represented either by Object Detection Techniques (ODT) or by raw images. The agents were evaluated in terms of the rate of goals obtained in the free-kicks scenarios provided by FIFA. The results show that DAI performs well on both the single and multi-agent based platforms. Further, the raw image based technique proved to be better than the ODT based method to represent the environment.