Deep Reinforcement Learning to train agents in a multiplayer First Person Shooter: some preliminary results

摘要

Training agents to play in contemporary multiplayer actions game is a challenging task, especially when agents are expected to cooperate in a hostile environment while performing several different actions at the same time. Nonetheless, this topic is assuming a growing importance due to the rampaging diffusion of this game genre and its related e-sports. Agents playing in a multiplayer survival first person shooter game should mimic a human player, hence they should learn how to: survive in unexplored environment, improve their combat skills, deal with unexpected events, coordinate with allies and reach a good ranking among the players community. Our aim has been to design, develop and test a preliminary solution that exploits Proximal Policy Optimization algorithms to train agents without the need of a human expert, with the final goal of creating teams composed only by artificial players.