An Evolutionary Game Tree Search Algorithm of Military Chess Game Based on Neural Value Network

Incomplete information game is an important research direction in artificial intelligence. Military chess is a typical incomplete information game. Since a player does not know the opponent's chess type in the whole game, military chess has high requirements for the design of the situation evaluation algorithm and the search algorithm. Most of the current military chess AI models are based on human experience. However, such models have disadvantages, such as uncertainty and the difficulty in adjusting parameters. In this work, a three-level decision model for incomplete situation is presented. Opponent imitation and feature extraction of the situation are promoted based on game rules while the neural network and the situation evaluation model are combined to help searching the game tree. Chess manual and self-play games are used to train the model. Our proposed method is verified with experiments of playing against the AI model from the National Computer Game Competition.