Research On 2-on-2 Three Kingdoms Reinforcement Learning Algorithm And Platform Implementatio

Multi-agent reinforcement learning is an important problem in artificial intelligence research,and breakthroughs have been achieved in combining game theory in typical card games such as Go and Texas Hold’em.Three kingdoms killing game is a popular card game with real-time strategy and cooperative competition,and its game environment can be used to study reinforcement learning algorithms to better learn and make intelligent decisions in the battle process.However,the research on Three Kingdoms Killing games is now limited to game development,and the problems in the field of multi-agent limit the application of this method in three kingdoms killing games.This thesis designs a 2-vs-2 three kingdom killing game environment,and based on the reinforcement learning method,introduce a value decomposition multi-agent reinforcement learning algorithm for effective signal decision-making and sparse reward problems in the field of multi-agents,and finally implements the 2-vs-2 three kingdom killing game battle platform.The main work is as follows:(1)SGS-VDMA for 2-vs-2 three kingdom killing multi agent algorithm based on value decomposition.In the multi-agent environment 2-vs-2 three kingdom killing game,how to model the actions and states of the four agents,so that the agents can communicate effectively,and to solve the sparse reward problem is a difficult point in designing the algorithm.In order to solve the above problems,this thesis proposes a reinforcement learning algorithm SGS-VDMA,which is based on the Actor-Critic framework for network design.The algorithm changes the action value evaluation to state value evaluation,makes multiple agents make decisions in the environment by decomposing the state value,and generates internal rewards based on the cooperation willingness between agents.Besides,it solves the problem of sparse rewards.According to the basic rules of the game,a 2-vs-2 three-kingdom killing game environment is constructed,and the multi-agent reinforcement learning algorithm SGSVDMA is implemented in the environment for agent decision-making in the same camp,and compared with several commonly used reinforcement learning game algorithms COMA,MAPG,Single＿AC and random method used by adversary camp,the results show that this method is superior to the other decision-making methods: after multiple training and testing,the camp that uses the SGS-VDMA algorithm receives higher team reward values and have a win rate of at least 6% higher than the camp that uses other decision-making methods.(2)Based on web development and other technologies,the 2-vs-2 three kingdoms killing game battle platform has been realized.Based on web development and other technologies,the 2-vs-2 three kingdoms killing game battle platform is designed and implemented.Users can conduct human-machine,algorithm(machine-machine)battles in visual scenarios,and the platform provides a simple interface for scholars to upload their algorithms,to train and to test in the game environment of the platform.Besides,users can manage the reinforcement learning algorithms in the local algorithm library and the data generated by battle training.