Reinforcement learning compared to rule-based play in UNO

Sammanfattning

This thesis investigates the application of reinforcement learning (RL) algorithms in playing the card game UNO compared with traditional rule-based approaches. UNO, with its strategic complexity and dynamic gameplay, presents a rich domain for exploring the adaptability and effectiveness of various AI strategies. By employing the RLCard library, an open-source platform designed for developing and benchmarking RL agents in card games, we systematically compared the performance of selected RL algorithms—Deep Q-learning (DQN), Neural Fictitious Self-Play (NFSP) and Deep Monte-Carlo (DMC)—against a rule-based benchmark. Our research methodology encompasses the development and training of RL agents, followed by a comprehensive analysis of their gameplay performance compared to rule-based agents. Through empirical research, we quantified the agents’ success rates and efficiency in learning optimal strategies. The results of our study indicate significant performance differences among the tested algorithms. While the NFSP agent struggled with effective learning and consistently lost to both random and rule-based agents, the DQN agent showed incremental improvement before plateauing. The DMC algorithm emerged as the top performer, significantly outperforming the other models. Overall, the DMC model's superior performance underscores its potential for future research.