Research On Incomplete Information Game Based On Improved DenseNet And CNN

In recent years,artificial intelligence technology has developed rapidly.As an important branch of artificial intelligence technology,deep learning technology is widely used in image recognition,object detection and speech recognition.Since the Google team applied deep learning to complete information game go,and achieved great success,more and more research on the application of deep learning technology to the game began.Traditional methods to solve incomplete information game mainly include reinforcement learning and search.For the reinforcement learning method,for the problem with huge action space,it needs to maintain a huge action value table,so it is very difficult for many problems to maintain the table,and the computing power of the computer is also very high.When modeling the problem,the search method also needs to create a huge search tree due to the huge action space,which is acceptable for the complete information game with small action space.However,the problem of huge state action space has great limitations.This paper takes incomplete information game as the research object,aiming at the limitations of traditional methods to solve incomplete information game,and uses the powerful feature extraction ability of deep learning method to solve the decision-making problem in mahjong game directly by end-to-end method.The content of this paper is composed of the following parts,each of which represents a brief overview of the work I have done in this subject.1、This paper designs a three-dimensional matrix which can be used to describe the game information of incomplete information game scenes,in which the number of channels represents each influencing factor.For this topic,the training data set is from the game company’s game player log information.The processing of log information includes a series of data processing operations,such as ranking,data filtering,data extraction and so on,so that the training set data can meet the input data structure requirements of neural network model,that is,from log information format to three-dimensional matrix format.By using the above data construction method,not only can effectively reduce the calculation amount of the traditional method of constructing game tree to solve decision problems,but also the method of deep learning is more simple and convenient to extend to other decision problems.2、This paper designs a mahjong playing decision model based on improved densenet and an improved CNN action network model.Considering the problems of gradient vanishing and over fitting of deep neural network model due to the increase of model depth.This paper introduces the fire module of squeezenet network model into densenet network model.Using the data compression function of fire module and the design of hybrid convolution kernel structure,the multi-level feature extraction of 3D matrix data is carried out,which not only reduces the complexity of the model,but also effectively alleviates the over fitting of the model by the way of dense connection of the original model.For the CNN network model,inspired by the Google net network model,this paper improves the CNN model by combining the concept structure with the hop layer idea of residual network.We train and test the improved densenet network model and the improved CNN network model on the same data set.Compared with the classical machine learning model and deep learning model,the improved model can improve the accuracy,The purpose of the experiment was achieved...