Title: Training a popular Mahjong agent with CNN and self-attention

Authors: Liu Liu; XiaoChuan Zhang; ZeYa He; Jie Liu

Addresses: School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401120, China ' School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401120, China ' School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401120, China ' School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401120, China

Abstract: Popular Mahjong is a variety of Chinese Mahjong that is characterised by having a high likelihood of Chow and Pong. To search for the hidden information in the Popular Mahjong time series decision-making data and arrive at reasonable Discard, Chow and Pong decisions. In this paper, a hybrid decision model that combines convolutional neural network, long short term memory and self-attention mechanism is proposed. A Mahjong agent JongMaster is created in this essay. Five kinds of action models are created by fusing the aforementioned hybrid model with real-world knowledge: Discard, Chow, Pong,Kong, and Riichi. These models work together to create the JongMaster's decision-making process. Lastly, JongMaster and the three benchmark agents conducted 1000 rounds of combat tests on the professional test platform. JongMaster increased the Hu rate and the highest score of a single round by 10.5% and 9 points, respectively, and decreased the shooting rate by 5.12%.

Keywords: popular Mahjong; deep convolution network; self-attention mechanism; LSTM; long short term memory.

DOI: 10.1504/IJCSM.2024.137266

International Journal of Computing Science and Mathematics, 2024 Vol.19 No.2, pp.157 - 166

Received: 18 Jul 2022
Accepted: 29 May 2023
Published online: 08 Mar 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article