Design and Implementation of an Einstein Chess Intelligent Gaming System Based on Policy-Value Network

Gengyun He; Gang Wang

doi:10.54097/g2e7vb65

Authors

Gengyun He
Gang Wang

DOI:

https://doi.org/10.54097/g2e7vb65

Keywords:

Einstein Chess, Policy-Value Network, Monte Carlo Tree Search, Deep Reinforcement Learning, PyTorch

Abstract

Einstein Chess is a board game characterized by randomness and special movement rules, presenting new challenges for artificial intelligence systems due to its complex game mechanics. This paper proposes an intelligent game system for Einstein Chess based on a policy-value network, effectively addressing randomness and special rules through an improved AlphaZero framework. The main innovations of the system include: (1) the design of a 16-channel state representation method that effectively encodes the board state and move history; (2) the construction of a deep convolutional network based on the PyTorch framework, utilizing shared feature extraction layers and a dual-head output structure; (3) the optimization of randomness response strategies through an improved Monte Carlo Tree Search algorithm. Experimental results indicate that the policy-value network-based intelligent game system for Einstein Chess achieved a significant advantage compared to the champion program of the 2024 Liaoning Province Computer Game Competition, attaining a win rate of 82.4% over 500 games. Through self-play training, the system demonstrated good learning capabilities and effective strategy improvement. This research provides a new solution for handling board games with randomness and offers a reference framework for the development of similar game systems.

Downloads

Download data is not yet available.

References

[1] CAI Biao, XU Xinyi, XIE Ting, et al. Research on improved deep neural networks in Einstein Chess[J]. Journal of Chongqing University of Technology(Natural Science), 2024, 38 (05): 108-114.

[2] XUE Yonghong, WANG Hongpeng. History and enlightenment of computer chess: From Deep Blue to AlphaZero [J]. Science & Technology Review, 2019, 37(19): 87-96.

[3] Silver D, Huang A, Maddison C, et al. Mastering the Game of Go with Deep Neural Networks and Tree Search [J]. Nature,2016, 529 (7587): 484 - 489.

[4] Silver D, Schrittwieser J, Siomonyan K,et al. Mastering the Game of Go without Human Knowledge [J]. Nature,2017,550 (7676): 354 - 359.

[5] ZHOU Zilong. Implementation and optimization of game search tree algorithms[J]. Scientific and Technological Innovation, 2021, (18): 108-110.

[6] MING Hongbing. Design and implementation of deep neural networks for Monte Carlo tree search algorithm[D]. Northeastern University, 2021. DOI:10.27007/d. cnki.gdbeu. 2021. 001565.

[7] GAO Tongtong, DING Jiahui, SHU Wenao, et al. Research on NoGo game system based on AlphaZero[J]. Intelligent Computer and Applications, 2022, 12(11): 138-141+147.

[8] ZHANG Zeyang. Research and implementation of perfect information game theory based on reinforcement learning[D]. Xidian University, 2021. DOI:10.27389/ d.cnki.gxadu. 2021. 002956.