Play games using Reinforcement Learning and Artificial Neural Networks with Experience Replay

• Published in: 2018 IEEE ACIS 17th International Conference on Computer and Information Science (ICIS) pp. 855 - 859
• Main Authors: Xu, Meng, Shi, Haobin, Wang, Yao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2018
• Subjects: Approximation algorithms; artificial neural network; Artificial neural networks; Biological neural networks; Convergence; experience replay; Games; Learning (artificial intelligence); Non Player Character; reinforcement learning
• DOI: 10.1109/ICIS.2018.8466428

Reinforcement learning is a self-learning algorithm in which an agent acquires experience through continuous interaction with the environment, which is more like the process of human or animal learning. Reinforcement learning is widely used in the field of playing games, and the classical reinforcement learning algorithm can easily produce Curse of Dimensionality when the state dimension is too large. In order to improve the convergence rate of reinforcement learning, a method of training Non Player Character (NPC) in games using Sarsa learning algorithm is proposed. The artificial neural network is used to approximate the value function. In order to make better use of experience, this paper sets up double neural networks, and uses experience memory to store experience, and uses experience replay to speed up the convergence of sarsa learning. Using the method presented in this paper to train NPC, we can find the NPC which is trained by the method has more learning ability than the classical reinforcement learning.