In some sense, computer game can be used as a test bed of artificial intelligence to develop intelligent algorithms. The paper proposed a kind of intelligent method: a reinforcement learning model based on temporal difference (TD) algorithm. And then the method is used to improve the playing power of the computer game of a special kind of chess. JIU chess, also called Tibetan Go chess, is mainly played in places where Tibetan tribes gather. Its play process is divided two sequential stages: preparation and battle. The layout at preparation is vital for the successive battle, even for the final winning. Studies on Tibetan JIU chess have focused on Bayesian network based pattern extraction and chess shape based strategy, which do not perform well. To address the low chess power of JIU chess from the view of artificial intelligence, we developed a reinforcement learning model based on temporal difference (TD) algorithm for the preparation stage of JIU. First, the search range was limited within a 6×6 area at the center of the chessboard, and the TD learning architecture was combined with chess shapes to construct an intelligent environmental feedback system. Second, optimal state transition strategies were obtained by self-play. In addition, the results of the reinforcement learning model were output as SGF files, which act as a pattern library for the battle stage. The experimental results demonstrate that this reinforcement learning model can effectively improve the playing strength of JIU program and outperform the other methods.

In this study, hybrid state-action-reward-state-action (SARSAλ) and Q-learning algorithms are applied to different stages of an upper confidence bound applied to tree search for Tibetan Jiu chess. Q-learning is also used to update all the nodes on the search path when each game ends. A learning strategy that uses SARSAλ and Q-learning algorithms combining domain knowledge for a feedback function for layout and battle stages is proposed. An improved deep neural network based on ResNet18 is used for self-play training. Experimental results show that hybrid online and offline reinforcement learning with a deep neural network can improve the game program’s learning efficiency and understanding ability for Tibetan Jiu chess.

Go Artificial Intellects(AIs) using deep reinforcement learning and neural networks have achieved superhuman performance, but they rely on powerful computing resources. They are not applicable to low-cost personal computer(PC). In our life, most entertainment programs of Go run on the general PC. A human Go master consider different strategies for different stages, especially for the middle stage that has a significant impact on winning or losing. To study arguably a more humanlike approach that is applicable to low-cost PC while not reducing chess power, this paper proposes a new search algorithm based on hypothesis testing and dynamic randomization for the middle stage of the game Go. Firstly, a new method to decide the intervals of different playing stages more reasonable based on hypothesis testing is proposed. Secondly, a new search algorithm including a layered pruning branch method, a comprehensive evaluation function and a new selecting node method is proposed. The pruning method based on domain knowledge and upper confidence bound formula(UCB) are all applied to subtract the branches from the lower evaluation score, which was ranked behind 20%. The comprehensive evaluation function with adjustable parameters is proposed to evaluate the tree nodes after pruning. The new selecting node method based on dynamic randomization is used to expand the tree by selecting a node randomly from the high-quality node interval. Finally, the experimental results show that the designed algorithm outperforms Gnugo3.6 and Gnugo3.8 in chess power while reducing average search time and average RAM cost for one move effectively on a 19×19 board.

Computer game-playing programs based on deep reinforcement learning have surpassed the performance of even the best human players. However, the huge analysis space of such neural networks and their numerous parameters require extensive computing power. Hence, in this study, we aimed to increase the network learning efficiency by modifying the neural network structure, which should reduce the number of learning iterations and the required computing power. A convolutional neural network with a maximum-average-out (MAO) unit structure based on piecewise function thinking is proposed, through which features can be effectively learned and the expression ability of hidden layer features can be enhanced. To verify the performance of the MAO structure, we compared it with the ResNet18 network by applying them both to the framework of AlphaGo Zero, which was developed for playing the game Go. The two network structures were trained from scratch using a low-cost server environment. MAO unit won eight out of ten games against the ResNet18 network. The superior performance of the MAO unit compared with the ResNet18 network is significant for the further development of game algorithms that require less computing power than those currently in use.

With the development of deep learning and graph deep learning, the network structure is more and more complex, and the parameters in the network model and the computing resources and storage resources required are increasing. The lightweight design and optimization of the network structure is conducive to reducing the required computing resources and storage resources, reducing the requirements of the network model on the computing environment, increasing its scope of application, reducing the consumption of energy in computing, and is conducive to environmental protection. The contribution of this paper is that Geometry-V-Sub is a graph learning structure based on spatial geometry, which can greatly reduce the parameter requirements and only lose a little accuracy. The number of parameters is only 13.05–16.26% of baseline model, and the accuracy of Cora, Citeseer and PubMed is max to 80.4%, 68% and 81.8%, respectively. When the number of parameters is only 12.01% of baseline model, F1 score is max to 98.4.