Investigating Reinforcement Learning for Dynamic Difficulty Adjustment

Dynamic Difficulty Adjustment (DDA) is a technique to automatically adjust various game factors, such as items, maps, or opponent behavior, to provide players with a challenging and engaging experience. The goal is to maintain a balance ensuring an optimal level of enjoyment. In this work, we propose a reinforcement learning agent in a fighting game to create an opponent that matches the player's skill level. We propose a reward function that leads the player to have similar relative skill to his opponent and maintain a balanced match. Additionally, we introduce a penalty given to the agent during training to constrain its win rate. Therefore, creating an opponent that is not too wear nor too strong. We also explore regularization techniques to improve the agent's performance and adaptability. We show that regularization improves over the baseline in generalizing its behavior to handle opponents not encountered during training.