Research On Hybrid Intelligent Method Oriented To Human-Machine Sequential Decision Making

With the development of artificial intelligence technology,machine intelligence has been continuously improved,followed by the application and development of machine intelligence in all walks of life.In this process,it is inevitable that the autonomy of the machine is not enough to solve the situation that it should be solved by humans or that humans must participate in decision-making.It is particularly important and meaningful to consider the decision-making problem of human intelligence and machine intelligence in this scenario.More specifically,as a kind of sequential and multi-stage dynamic decision-making problems,the development of sequential decision-making problems is closely related to the fields of engineering applications,production and life in the current artificial intelligence era.The role of humans are reflected in two aspects of sequential decision-making problems.One is that humans themselves are part of the sequential decision-making problem model,that is,such problems are inseparable from humans,such as minimally invasive surgery.Second,humans' relevant information is not reflected in the sequential decision-making problem model,but because of the unique cognitive ability of humans,it can appear in the problem-solving method to achieve the purpose of improving the problem-solving,such as in the machine search and rescue system.We collectively refer to the above two scenarios as the"humanmachine sequential decision-making problem".For the human-machine sequential decision-making problem,due to the essential difference between human intelligence and machine intelligence,and the huge difference in mathematical expression,when humans and machines work together to solve the problem,it is inevitable that the quality of decision-making is not high or even the phenomenon of decision-making errors due to coordination reasons.However,the direct application of the control algorithm of the traditional human-machine system cannot effectively deal with these problems,which causes the failure of the machine agent,the waste of manpower,and even the performance deterioration or even the collapse of the decision-making system.Therefore,it is urgent to design an effective humanmachine hybrid intelligent algorithm to solve these problems.This thesis takes the human-machine sequential decision-making problem as the research object,and researches on three issues:the division of decision-making authority in human-machine hybrid intelligent control,the timing of triggering switching of traded control and the degree of mixing of shared control.It aims to propose an effective human-machine hybrid intelligent decision-making algorithm to improve the solution of the problem of improving the human-machine sequential decision-making.The main content of this thesis mainly includes the following aspects:1.A human-machine hybrid intelligent control framework based on reinforcement learning method is proposed.By arbitrating decisions made by machine agents and humans with credibility and safety as evaluation indicators,the decisionmaking actions to be executed are determined more optimally.Considering the model-based reinforcement learning subsystem and the model-free reinforcement learning subsystem,it provides more possibilities for adapting to a wide range of sequential decision-making application scenarios.2.Considering the traded control in the human-machine sequential decision-making problem,the concept of autonomy and autonomy boundary is proposed.By formalizing the solution of the autonomous boundary as a conventional optimization problem related to the task goal for discussion and judgment,the control scheme and algorithm of the intervention control are optimized.And the decision performance of human intervention machine and machine intervention human in the human-machine sequential decision-making process is improved.3.Considering the shared control problem in the human-machine sequential decision-making problem,a hybrid parameter optimization design scheme based on the autonomous boundary is proposed,which directly affects the generation of the final action to be executed by adaptively adjusting the hybrid parameter.Taking into account the degree of integration of human and machine actions,the optimal solution appears in the expanded space formed by the human action space and the machine action space,which provides more space for the improvement of decision-making quality.4.Considering the inaccurate problem of the single value estimation of the autonomous boundary in traded control and shared control,an uncertainty estimation method based on Bayesian neural network is proposed to obtain the probability distribution information of the autonomous boundary and use it for decisionmaking action generation.By using the uncertainty of the autonomous boundary to optimize the design of the human-machine hybrid intelligent algorithm,it not only makes more choices for the optimization of decision-making actions,but also more in line with humans' vague thinking about the decision-making boundary.In summary,this thesis systematically studies the problems faced by the humanmachine hybrid intelligent algorithm to solve the human-machine sequential decisionmaking,innovatively proposed corresponding solutions,and greatly promoted the human-machine sequential decision-making solution.