Research And Application Of Incomplete Information Game Decision Method Based On Deep Learning

Human beings are making tradeoffs and decisions all the time in their daily life.These problem-solving scenarios can be abstracted into game decision-making matters.Game problems can be divided into two categories: the complete information game and the incomplete information game.The problem of incomplete information game usually refers to the process in which the players can not fully grasp all the information in the game process.In real life,such as commercial negotiation,information security,advertising pricing,military deduction,game entertainment as well as many other issues can be summarized as incomplete information game.With the development of artificial intelligence and the increasingly widespread of deep learning applications,the problem of incomplete information game decision making through the deep neural network has become the focus in research of machine game which also has much practical significance.At present,the traditional method to solve the incomplete information game problem is to model the game problem with the help of reinforcement learning,design the reward function and construct the game tree to transform the game problem.The final step is to optimize the state-action value function of the game strategy with the help of the game tree search and value iteration.The traditional method can perform very well in the complete information game and the small-scale incomplete information game.However,in the face of the complex incomplete information game,it doesn’t guarantee the value function converges.Also,the game tree is too large to simulate,and the model training cost is too expensive in that situation.To solve these problems,in this paper a method to solve incomplete information game decision-making problems under complex background by using the deep neural network is proposed.The main researched work includes the following parts:1.A semantic segmentation method based on knowledge and rules is designed to model the game problem and extract the observable information of incomplete information game with other important factors related to the game decision.Moreover,this information is compressed into a three-dimensional multi-channel image.Semantic segmentation method with game rules as the basic element,which is giving a complete description of the current situation of significant information.Moreover,the based on simple game knowledge,the key factors that may affect decision making are constructed into low-level images.In this way,the complex calculation process of building a game tree is replaced by the simpler one.And it is more friendly to train the deep neural network in this form of data.2.The method of incomplete information game decision model training based on an improved deep residual network is designed.Considering the deep learning model with neural network layer deepening may cause problems such as the gradient disappearing.In this paper,our innovative network structure,different from the original residual network topology,is combining multi-scale asymmetric convolution for multilevel feature extraction of image information,as well as recycling more parallel branch network structure to increase width to enhance model learning ability.With the combination of multiple new substructures named Inception+ and the shortcuts connection to implement the residual network identity mapping,a new residual block named GoBlock structure is proposed.Experiments show that,based on the same game scene image dataset,our improved deep residual network is more accurate than other traditional machine learning methods and deep learning methods in the prediction of incomplete information game decision classification problem.3.An intelligent decision-making system of incomplete information game based on deep learning is designed.In a multiplayer incomplete information game problem(competitive mahjong game)under a complex background,an intelligent decisionmaking system is realized by training deep neural network model with real battle data.The system played against decision systems generated by other machine learning algorithms.At the same time,the system is deployed in an online competitive mahjong game application against real human players.Experiments show that the intelligent decision system proposed in this paper has a better performance than other decision systems in terms of average score and average win rate.In 5,900 games,the intelligent decision system has a higher average win rate than normal human players.The analysis of battle replays further prove that our intelligent decision-making system has a high level of game decision-making intelligence.