A Human-centered Human-machine Function Allocation Method Based on Reinforcement Learning

Human-machine function allocation or adaptive automation is an important means to achieve human-machine integration in human-machine collaboration systems. One of the purposes of human-machine function allocation methods is to maintain the operator's workload state in a reasonable state while realizing human-machine function allocation. Inspired by previous studies that widely used reinforcement learning methods for machine system control, this study modeled the automation level adjustment events in the human-machine interaction process using Markov decision processes. Taking the operator's physiological state that can be collected in real time during the human-machine interaction process as the model input and the automation level adjustment action as the output, a human-centered automation level adjustment model was established using the deep Q-network method. In addition, this study carried out a human-machine interaction experiment based on the flight control scenario to obtain data for training and validating the model. The results show that the model proposed in this study can better adjust the automation level of the system to maintain and balance the operator's workload state.