Research On Game Algorithm Of Imperfect Information 3D Video Game Based On Deep Reinforcement Learning

Artificial intelligence research is in full swing around the world,and machine game has been a hot field of artificial intelligence research since the birth of computer and game theory.In recent years,the success of AlphaGo Agent has attracted more researchers who devote time to the research of machine game.In the game research,it is worth noting that the deep learning and reinforcement learning algorithms mainly used by the agent have become the most representative technology in the current artificial intelligence wave.The success of Go Agent represents a historic breakthrough in the complete information game in the field of machine games,and there are still many topics that need to be studied and solved in the incomplete information game with higher complexity which is closer to the real world.On the other hand,the game has become the most important research tool and testing platform in the field of artificial intelligence,especially machine game.Because it has clear definition of rules,rich scenes and reusability.In this paper,the incomplete information 3D video game is selected as the carrier of the research content.For the characteristics of the state dimension and the action space in the incomplete information game,the deep neural network is used to represent the value function and strategy model in the reinforcement learning,and the original game screen is directly used.As the input of the neural network,it is different from the traditional reinforcement learning algorithm,which requires the use of relevant domain knowledge to solve the strategy through task modeling and manual extraction of features.Therefore,the problems of poor scalability,low efficiency,and inability to converge are solved.Aiming at the shortcomings of the high variance of the original strategy gradient algorithm in the reinforcement learning algorithm,this paper introduces the setting of the baseline function.At the same time that the baseline function is optimized,the value model is introduced,and the strategy gradient algorithm that is added with value model is proposed.The problem of narrow scope of application and the fitting difficulty of the reinforcement learning algorithm based on value function is solved.In order to improve the sampling and training speed of the reinforcement learning algorithm,a parallel training mechanism is added.Considering that the agent needs to comprehensively consider and analyze the characteristics of history and current information,the gate structure is used to add the memory unit to the deep reinforcement learning neural network based on the value model and the strategy gradient model.In order to solve the problem of sparse reward that often occurs in reinforcement learning tasks with high-dimensional space,the reward value is supplemented by reward design and self-driven mechanism which at the same time makes the intelligent body fully explores the environment.Through the target detection technology,the deep reinforcement learning network is provided with the location information of the enemy in the current game screen which are useful features.Using ViZDoom as the test platform of deep reinforcement learning algorithm,the relevant experimental analysis and competition results show that these improved algorithms can enhance the game level of the agent.In order to help the game agent improve the level and improve the efficiency of implementing or enhancing the reinforcement learning algorithm,this paper designs and implements a set of agent level analysis system.