Research On Agent Game Problem Based On Reinforcement Learning

With the development of unmanned technology,how to better use unmanned equipment has become an urgent problem to be solved.And with nobody to the agent game problems in application of technology of related study,the classical method mostly lack of model abstraction,often need to be done in the process of problem solving more human intervention,and enhance the learning interaction with questions,automatic modeling and solving the problems,but current related research,enhance the learning areas more is based on a simple question research the improvement to the algorithm,and the characteristics of the actual problem in large.Therefore,the research objective of this paper is to apply and improve the existing mainstream reinforcement learning algorithm in the actual agent game problem,so as to solve the game problem to some extent.Firstly,this paper analyzes and models the problem of agent game.At the same time,in order to facilitate the algorithm design of reinforcement learning,mathematical models of decoupled path planning problem and attackdefense-counterattack problem are established according to the decoupling principle.Then,based on the above mathematical model,the simulation environment of agent game problem is built to interact with the augmented learning algorithm.In the part of algorithm research,the state design and reward design were firstly studied for the problem of path planning and attack-defense confrontation,and the mainstream strategy gradient,DQN and A2 C reinforcement learning algorithms were applied to solve the problem.Combined with the learning effect and problem analysis,the network structure was improved into parallel network structure and branch network structure.From two perspectives of different algorithms and different improvement methods,the results of the algorithm are compared to evaluate the branch A2 C method which is most suitable for the original problem.Finally,based on the study of path planning and defense against problems as a result,the design condition and reward of the original problem,application of the above strategies gradient,DQN and A2 C enhance learning algorithm of normal network structure,network structure and parallel structure,branch of intelligent game problem solving,and verify the effectiveness of the decoupling problem algorithm based on the analysis results.The parallel network structure designed USES different networks to make decisions of the decoupled actions in the game problem,which reduces the difficulty of each independent network learning.Although it increases the instability of learning,it improves the learning effect significantly.The designed branch network will decouple the actions of each other,feature extraction layer of public network,and adopt different network output structure,which not only reduces the learning difficulty of the network,but also reduces the learning instability.Finally,the learning difficulty of agent game problem is reduced,and the learning ability of enhanced learning algorithm is expanded.