DEEP REINFORCEMENT LEARNING FRAMEWORK FOR PLAYING FIRST-PERSON SHOOTER OVERWATCH 2

Abstract

The application of deep reinforcement learning (DRL) to first-person shooter (FPS) video games offers a compelling avenue for advancing real-world applications, particularly in domains requiring autonomous navigation and decision-making in complex 3D environments. Autonomous vehicles, smart wheelchairs, and robots operating with limited information about their surroundings can benefit significantly from insights gained by training DRL agents to play FPS games, relying solely on visual input from a first-person perspective. DRL, which combines deep learning (DL) and reinforcement learning (RL), demonstrates success in various fields, including complex game playing. RL, a paradigm that teaches agents optimal behavior through reward functions and planning, achieves superhuman performance in games like Go and chess, owing to the well-defined rules and reward structures inherent in game environments. This thesis presents the development of a novel DRL framework for playing FPS games, specifically targeting the complex, hero-based multiplayer FPS game Overwatch 2. Leveraging computer vision (CV) from the first-person perspective, the research culminates in a robust agent, RLWatch, capable of playing a specific Overwatch 2 scenario at a performance level comparable to that of skilled human players. The game environment is established to closely mirror the state-of-the-art (SOTA) environment ViZDoom, while incorporating the complexities of Overwatch 2. SOTA DRL models, such as asynchronous-advantage-actor-critic-anticipator (A3C-Anticipator), serve as a foundational architecture for the framework. The results demonstrate the effectiveness of the RLWatch framework, showcasing its ability to achieve high performance in a complex multiplayer FPS environment.