Design And Implementation Of Game Decision-making Mechanism Based On Deep Reignfrocement Learning Algorithm

Game decision-making process is a key problem in artificial intelligence and plays a vital part in chess entertainment,business managing and military simulation.Traditional decision-making process is like walking in a game tree.Nodes in tree are states and edges came from a node are decisions under that state.Decision makes one state migrate to another.Simulating all states and decision made under those states using enormous computation with search tree pruning,transposition table and parallelization has made huge progress.However,top game decision-making program still can’t match for top human experts due to the massive search space.Progress in game decision-making using traditional searching algorithm has descended into stagnation,with little upside.Aim at above problems,this article has presented a game decision-making mechanism based on deep reinforcement learning.Game decision-making process can be abstracted to interaction between agent(decision making AI)and environment(state space).Main job of this article includes:1)Introduce deep reinforcement learning on policy generating and states evaluation of game decision-making,which improves fitting quality of policy and evaluation function.2)Design the system architecture of game decision-making mechanism based on deep reinforcement learning,including neural network module,self-play module and evaluator module,and modules are designed with low coupling degree and concise realization.3)Realize a game decision-making system based on deep reinforcement learning.Besides,we analyze function and performance of the system.The game decision-making engine we harvest is fully functional and powerful.Finally,in free opening mode,this gomoku decision-making engine obtains the wining rate of 90 percent in master area of QQ Gomoku and defeats YiXin by the wining rate of 55 percent.