Research And Application Of Incomplete Information Game Based On Search Tree And Opponent’s Card Strength Evaluation

For a long time,game problems have been the focus of artificial intelligence research.With the rapid development of artificial intelligence technology,many complete information game fields(such as Go,chess,etc.)have achieved outstanding results through machine game technology.Currently,the game of incomplete information is the focus of game research.Unlike complete information games,not all game information can be observed in incomplete information games,and the game state space is huge.Traditional full search trees cannot be completely expanded in an effective amount of time.Therefore,reasonable pruning operations are needed to reduce search time and improve search accuracy.This paper focuses on the Mahjong game problem in the field of incomplete information games,and uses heuristic information to guide the pruning of the game search tree.This reduces search time while also improving the game win rate.Deep learning technology is also used to compare and evaluate opponent hand strength,and the evaluation results are used to assist in search tree decision-making.The specific research work is as follows:1.To solve the problem of incomplete expansion of game search trees in Mahjong,heuristic information is used to guide the search direction based on the traditional complete search tree.Reasonable pruning is performed,and the game mode is simplified by reducing the search nodes from two types,discarding and drawing,to one type of drawing.This improves search efficiency and accuracy.Based on the pruning strategy,two different search modes are designed,which are shortest path search and deep exploration search,corresponding to low-risk low-yield strategies and high-risk high-yield strategies in the game scenario,respectively.Different expansion paths and evaluation functions are used to achieve appropriate decision-making in different scenarios in Mahjong game.2.A Mahjong feature coding method suitable for deep learning is used to solve the feature coding problem in this incomplete information game scenario.Unlike traditional opponent modeling that directly predicts opponent hands,this paper uses deep learning technology to compare and evaluate the strength of other players’ hands with their own to assist decision-making.Inception-ResNet-v2 is used as the backbone network,and ECA attention mechanism and feature fusion mechanism are introduced to enhance the model’s feature extraction ability.Game data under complete information is used to ensure the quality of the training data,and the evaluation accuracy is ultimately improved.3.The results of the evaluation of opponent hand strength are integrated into the game tree to assist the Mahjong AI agent in making discarding decisions.The experiments show that the game tree decision-making AI agent with integrated opponent hand strength evaluation will adjust its search strategy according to the game scenario and is superior to the search tree decision-making AI agent with a single search mode and the Mahjong AI agent processed by other methods such as rules.