Playing First-Person-Shooter Games with A3C-Anticipator Network Based Agents Using Reinforcement Learning

Common built-in bots act upon pre-written scripts to make decisions and actions where sometimes they acquire and take advantage of unfair information, instead of acting flexibly like human players, who make decisions only based on game screens. This paper mainly focuses on studying the applications of deep learning and reinforcement learning in the field of computer game agents. The goal is to create agents that make decisions in human’s way and gets rid of relying on unfair information. A game agent is implemented in line to the A3C algorithm. This agent takes the original real-time game screen as an input to the network and then outputs the corresponding discrete actions. The agent can interact with Viz and read the real-time game screen to make decisions for controlling the characters. This paper made an improvement to the A3C algorithm by adding an anticipator network to the original model structure. The goal of doing this is to make the agent act more like human players. It generates anticipation before making decisions, and then combines the real-time game screen with anticipation images together as an input to the network defined by the A3C algorithm. The result shows, that the A3C algorithm with Anticipation performs better than the original A3C algorithmd.