Research And Implementation Of Computer Game Algorithm Of "Jiuqi

Tibetan Jiu Chess is a complex,diverse,and unique chess sport that effectively simulates the process of strategic thinking between players,making it an ideal subject for studying computer game-playing algorithms.Jiu Chess research is highly challenging and difficult.Current algorithms and models for Jiu Chess include pattern-based algorithms,temporal difference reinforcement learning algorithms,and deep reinforcement learning models that do not rely on expert knowledge.However,due to the limited number of game records,a lack of expert knowledge,and insufficient computing resources in ordinary laboratories to support self-play deep reinforcement learning model training,the playing strength of various game-playing agents has not yet reached an ideal level.To address these issues and develop a Jiu Chess-playing agent program that closely matches the playing strength of top-level players while promoting Tibetan Jiu Chess culture,the author conducted field research in Lhasa,Tibet,gaining an in-depth understanding of Jiu Chess strategies and conducting research on computer game-playing algorithms for Jiu Chess.The author established a phased Tibetan Jiu Chess expert knowledge base and developed a game-playing program based on expert knowledge to verify the effectiveness of the expert knowledge.The author also researched and developed a Jiu Chess game-playing program that combines expert knowledge and self-play deep reinforcement learning to improve playing strength and created a Jiu Chess game platform to collect game record data,as detailed below.1.By playing against high-ranking Tibetan Jiu Chess players,a large number of game strategies were summarized and a Jiu Chess expert knowledge base was established.Based on the expert knowledge,a Jiu Chess game-playing algorithm was proposed.Through field research in Lhasa,Tibet,a large amount of high-quality Jiu Chess game data was collected,and the thinking strategies of top-ranked 6-dan players were investigated.Seventy-two expert knowledge rules,such as prioritizing capturing "sentry" pieces and attacking weak positions in balanced situations,were summarized.To verify the effectiveness of this expert knowledge in improving the playing strength of Tibetan Jiu Chess,14x14 and 8x8 game-playing programs based on expert knowledge were developed.The 14x14 game-playing agent participated in the "National College Computer Game Contest Tibetan Jiu Chess Tournament" in 2021,demonstrating good algorithm performance and acceptable playing strength,winning a national-level second prize.This validated that strategies based on expert knowledge could be effectively applied to Jiu Chess games and improve the level of game-playing software.2.The AlphaZero program,through self-play learning in Go,Chess,and Shogi,defeated the most advanced programs in these three games.The author improved the AlphaZero algorithm and applied it to the opening phase of Tibetan Jiu Chess,researching a Jiu Chess game-playing algorithm that combined deep reinforcement learning with expert knowledge.The Jiu Chess convolutional neural network was trained based on self-play game records,predicting the agent’s moves during the opening phase.The Monte Carlo Tree Search was guided by these predictions to determine move choices.Through extensive selfplay training,the model gradually evolved into an agent with a considerable playing strength level during the opening phase.By combining this model with the move phase expert knowledge algorithm,a powerful and promising deep reinforcement learning model integrated with expert knowledge for Jiu Chess game-playing was developed.This model had a high win rate against purely expert knowledge-based models and achieved better performance than lowranking Jiu Chess human players in games against humans.3.The scarcity of Tibetan Jiu Chess game records is due to the lack of digitized game-playing platforms and the extremely low efficiency of manually collecting game records.The author developed an intelligent Jiu Chess gameplaying platform software using the Python language to collect game data.The software includes features such as human-computer play,human-human play,undoing moves,rule explanations,and automatic generation of game records.During the "Minshan Cup" National Tibetan Chess Open Tournament held in Lhasa,Tibet Autonomous Region in 2021,Jiu Chess masters from various provinces engaged in human-computer and human-human play using this software.The platform performed stably,collected high-quality game records,and greatly stimulated people’s interest in Jiu Chess games.