Using VizDoom Research Platform Scenarios for Benchmarking Reinforcement Learning Algorithms in First-Person Shooter Games

Abstract

Advances in deep reinforcement learning have made it possible to create artificial intelligence-based agents for games that use visual information to make decisions as accurately as humans. Novel procedures are often evaluated in two-dimensional games. However, they are relatively easy compared to three-dimensional games, which have a significantly larger state and action space and, more prominently, contain partially observable states. Thus, this paper trains agents with different reinforcement learning algorithms that work fine in contradiction of human players and in-built agents by evaluating them in the first-person shooter (FPS) game Doom using the VizDoom platform. The agents learned in three different scenarios (maps): ’ Defend the Center,’ ‘Deadly Corridor,’ and ‘Health gathering.’ C51-DDQN, DFP, and REINFORCE algorithms have been proven effective in this study. To assess how well the trained agents performed using various reinforcement learning algorithms, we compared the results of our research with other findings in the literature. Finally, this paper presents a comparative analysis and future research directions.

Keywords

• Artificial intelligence
• agents
• game AI
• reinforcement learning
• VizDoom