Deep stochastic weight assignment network of Chinese chess machine game

Man-machine game is an important component in the field of artificial intelligence. Game tree search algorithms and chess situation evaluation functions are mostly applied in the traditional chess game system. When the game tree method is used, the response time will be extended as the depth of tree. This paper proposes to use the stochastic weight assignment neural network (SWAN), trained by Extreme Learning Machine (ELM), to solve this problem. ELM is a fast learning algorithm in which all of the weights and biases between the input layer and hidden layer are randomized and the weights of output layer are analytically computed through solving a generalized inverse of matrix. Since the learning process does not include iteration, the learning speed is significantly accelerated and generalization capacity is improved greatly. Moreover, considering the complexity of Chinese chess situation features, feature learning and classification of chess samples cannot be efficiently accomplished by the shallow network. A deep stochastic weight assignment network (DSWAN) is developed to settle this difficulty.