Research On Multi-object Dynamic Decision-making For Cooperation And Confrontation Within Complex Environment

With the fast development of the artificial intelligence technology,multi-agent systems have been applied in many fields and the researches on multi-agent technology become more and more important.Furthermore,since the structure design technology of agents and the data acquisition technology have been more and more perfect,the reasonable situation assessment with the shared environmental data of the multi agents and the dynamic decision-making for cooperation and confrontation facing with the specific tasks have become important study points to improve the working ability and application value of the multi agents.For achieving the reasonable and efficient multi-agent situation assessment and decision-making for cooperation and confrontation,based on the multi-agent system researches,this paper proposed a completed improved multi-agent decision-making model for cooperation and confrontation.It is hoped that with this model,multi agents could perform the more reasonable and accurate information fusion and situation evaluation and then conduct the dynamic group self-decision-making.The content and innovation of this paper are listed as follows:(a)Design a situation assessment method based on deep learning algorithm.In order to realize the more reasonable and accurate situation assessment,complying with the uncertain correspondence relation from the environmental information to final situation assessment results,a situation assessment method based on deep learning algorithm is proposed: The normalized scene data is regarded as the input data of the input nodes of the deep neural network model and the situation labels are set as the outputs.The situation assessment network is trained following the training steps of deep learning.Finally,with the fuzzy theory,the situation assessment results are translated into the fuzzy situation vector.(b)Propose an intelligent decision-making method with the human inverse reinforcement learning algorithm.In order to improve the adaptability and the practicability of the decision-making system for the dynamic complex environment,with the obtained fuzzy situation results,an intelligent decision-making method with the human inverse reinforcement learning algorithm is presented in this paper: The fuzzy situation vector is taken as the state space,and determination domain of each state can be calculated by dividing the state space uniformly.Moreover,the decision schemes are regarded as the action set of the learning system.The human inverse reinforcement learning architecture is realized with the BP neural network method and the human reward mechanism.Finally,facing with the semi-Markov phenomenon,the cumulative reward updating strategy is improved to complete a whole group decision-making model for cooperation and confrontation based on the human inverse reinforcement learning.(c)Propose an optimization method based on fuzzy control theory for decision-making method with reinforcement learning.In order to improve the efficiency of decision-making method with the reinforcement learning and reduce the learning time cost,a dynamic learning rate method based on the fuzzy control is proposed: the influence degree of the action is taken as the input of the fuzzy control system and the learning rate value is set as the output.Through constructing the fuzzy inference engine,the input can be translated into the output with a smooth curve mapping and the learning efficiency can be promoted.Finally,the reasonability and the efficiency of proposed method are proved in the robot soccer cooperation and confrontation platform.