AN EVALUATION OF REINFORCEMENT LEARNING ALGORITHMS IN VIDEO GAME DEVELOPMENT

Reinforcement Learning (RL) has demonstrated substantial promise for creating adaptive, responsive AI in complex environments such as video games. Yet despite growing academic interest, industry adoption remains limited due to computational overhead, reward-design challenges, and unpredictable AI behaviors. This thesis investigates how RL algorithms—specifically Advantage Actor-Critic (A2C), Deep Q-Network (DQN), and Proximal Policy Optimization (PPO)—can be applied to three different genres of video games. Those being first-person shooter (fps), fighting, and strategy. Through a combination of scenario-based experimentation and comprehensive analysis, this work explores the feasibility and design considerations crucial for integrating RL-driven AI into commercial games. Key factors examined include environment complexity, reward shaping, and the tension between exploration and exploitation in dynamic gameplay settings. By evaluating each algorithm’s ability to adapt to partial observability and resource constraints, the thesis aims to uncover both the opportunities and the barriers that emerge when AI research collides with practical development needs. The main contribution of this study is a set of empirical insights and methodological guidelines that bridge the gap between academic RL research and real-world game production. These findings serve as a reference point for developers, researchers, and designers seeking to harness RL techniques effectively. By highlighting the unique challenges of each scenario and algorithm, the thesis lays the groundwork for future strategies that may bring RL from lab-based prototypes to fully realized, next-generation gaming experiences.