| With the progress of science and the continuous development of computer technology,the traditional multi-agent game antagonism method has been improved for a long time,and has been able to achieve relatively stable agent control in some simple environments.However,in the diverse,heterogeneous and dynamic environments such as military chess,the traditional game antagonism methods are too rigid and cannot respond quickly according to the situation,which directly leads to the poor performance of agents.In order to solve these problems,combining with the rapid development of artificial intelligence technology in recent years,this paper studies the multi-agent game confrontation based on reinforcement learning,and fundamentally changes the generation method of game strategy.Reinforcement learning,as one of the machine learning methods,does not need to directly consider the strategy choice of the game opponent,but only needs to let the agent interact with the environment and constantly optimize and update the control strategy through interactive data,which has many advantages in the field of game confrontation.In spite of this,reinforcement learning algorithm still faces many problems in the specific application of game confrontation,mainly manifested in: control task allocation design;Neural network structure design;Keep the stability of agent in the process of strategy iteration.Supported and funded by national Natural Science Foundation of China and Jiangsu Province Frontier Leading Technology Basic Research Special Project,this paper studies multi-agent game confrontation technology based on reinforcement learning on the basis of existing deep reinforcement learning and combined with specific complex game confrontation scenarios.The specific work and innovations are as follows:(1)to strengthen training in the process of game against environmental dynamic change,state space is larger,the training cycle is long,the opponent’s game against the unknown challenges,this paper adopts a take attention mechanism of proximal strategy optimization algorithm,based on the depth of the neural network to build game against agent,according to the preset reward function training,The problem of agent strategy generation in two-person zero-sum game scenarios is solved.By introducing attention mechanism innovatively,different information in the state space is given weight to shorten the training time and improve the training effect.Finally,this paper carries out simulation verification in a two-person zero-sum game scenario,and proves the feasibility of the algorithm by comparing with the training results of the traditional reinforcement learning algorithm.(2)for multi-agent game against agent quantity too much in the scene and environment partially observable challenge,this paper adopts a centralized training distributed execution of multi-agent reinforcement learning framework,based on the multi-agent environment QMIX algorithm game against strategy training,through the partially observable door control loop network state information processing,Finally,the concrete application problem of reinforcement learning algorithm in multi-agent game confrontation scenario is solved.In this paper,attention mechanism is innovatively added into the reinforcement learning training framework of centralized training distributed execution,and the high-dimensional state information in multi-agent games is screened to optimize the training process of reinforcement learning.Finally,the algorithm is applied and simulated in SC2 LE environment,and the reliability of the multi-agent reinforcement learning and training framework is proved through comparative experiments.(3)In view of the difficulties of complex tasks,diversified and heterogeneous combat units in the military game environment,this paper adopts the game strategy generation method of rule reinforcement learning joint decision making,and modularizes and decomposes the whole game confrontation decision task into independent modules according to the different task objectives.The reinforcement learning training based on subject-predicate action is combined with rule logic control to solve the problem that reinforcement learning is difficult to be directly applied in complex game confrontation scenarios.In this paper,we innovatively introduce the principal and predicate action space and the near end strategy optimization algorithm with attention mechanism into the strategy generation method,and realize the application of reinforcement learning algorithm in complex military scenarios by special processing of state space and action space in military game scenarios.Finally,the strategy generation method is applied and verified by simulation in military game scenarios,which proves the effectiveness of the strategy generation method. |
