Playing Oware Nam-nam with Deep Q-Networks

Abstract

In this thesis, deep Q-learning agents are used to learn how to play “Nam-Nam” Oware, a game known for its complex and unpredictable nature. We use Deep Q-Networks (DQN) and Double Deep Q-Networks (DDQN) in the study to simulate realistic gameplay. Techniques like self-play, epsilon-greedy strategy, and reward shaping are key to enhancing agent performance. The findings show that DDQN agents outperform DQN and random-play agents and produce more stable results. Our sensitivity analysis of hyperparameters further shows their influence on learning. This work not only advances reinforcement learning (RL) in strategic games but also demonstrates its potential to tackle real-world decision-making challenges that seem to be more random than strategic.