Browsing by Author "Marahemi, Sara"

Now showing 1 - 1 of 1

Open Access
Training agents to play cooperative games: A reinforcement learning approach
(Faculty of Graduate Studies and Research, University of Regina, 2023-07) Marahemi, Sara; Zilles, Sandra; Hamilton, Howard J.; Smith, Austen
Due to recent advances in research on reinforcement learning, self-learning agents are capable of accomplishing numerous kinds of tasks in complex environments without any prior knowledge. Recently, deep reinforcement learning algorithms have shown promising results in game-playing tasks that were previously impractical, including playing video games directly from raw screen pixels. In this project, we created a game engine for a card game called “The Mind”, and used reinforcement learning techniques in order to train agents to master this game. The Mind is a multi-player cooperative card game with the challenge of synchronizing the agents’ actions. We used Q-learning and deep Q-learning to estimate a Q-function which describes an agent’s best action to take at any state of the game. In this research, we implemented three types of agents based on two different reinforcement learning algorithms. The results showed that our trained agents performed better than random agent models. The highest testing win rate using the Q-learning algorithm was 86%. We also designed a reinforcement learning strategy, called observer learning, in which an agent updates its knowledge not only based on the feedback to its own actions, but also based on the feedback other agents receive for their actions. We reached the best testing win rates of nearly 99% for two Q-learning agents using our observer learning strategy in four levels of The Mind.