Deep Reinforcement Learning for Game Playing

Abstract

Deep Reinforcement Learning (DRL) has emerged as a powerful paradigm in the field of artificial intelligence. This paper explores the application of DRL in the context of game playing. We discuss how DRL algorithms, particularly Deep Q-Networks (DQN), have been used to achieve superhuman performance in a variety of games. The combination of neural networks and reinforcement learning has led to breakthroughs in game AI, allowing agents to learn complex strategies and adapt to dynamic environments.

This paper provides an overview of key concepts in DRL, including the use of experience replay, target networks, and the role of convolutional neural networks in processing game states. We also discuss the challenges and limitations of DRL, such as the need for substantial training data and the risk of overfitting.

To contextualize our discussion, we present a comprehensive review of relevant literature published before 2015. This includes seminal works by Watkins and Dayan (1992) on Q-learning, as well as notable contributions by Mnih et al. (2013) on DQN. We also delve into the history of reinforcement learning with temporal difference methods, citing Sutton's work (1988). Moreover, we highlight the significance of neuroevolution in game AI, drawing on Stanley's research (2002). Finally, we reference Barto and Sutton's textbook (1998) on Reinforcement Learning, which provides a foundational understanding of the subject.