Study On Multi-agent Consensus Based On Event-triggered Adaptive Dynamic Programming

Collaborative control has been applied in many engineering practice fields for the last several years,optimal consensus control for multi-agent system(MAS)becomes an current research topic.Designing the optimization protocol of consensus control in order to make the state of multiple agents converge under the protocol is the key to the research.There are many agents in the distributed system,which results in the complexity of the communication network structure.There are two main deficiencies in the existing theoretical research.Firstly,the conventional dynamic programming method will lead to dimension disaster when solving the optimization equation.Moreover,the analytical solution is troublesome to seek out for HJB equation.Secondly,for the research of existing consensus problem,the sampling and updating of the signal are carried out in a time-driven periodic sampling mode,which result in the decrease of network resources utilization.To overcome the above two deficiencies,event-triggered mechanism,dynamic programming,and reinforcement learning are combined.Based on it,the event-triggered adaptive dynamic programming is proposed to analysis the optimization problem of consensus control for several different MAS.The key points of this research are generalized as follows.(1)For the optimal consistency control problem of MAS with input constraints,event-triggered action-dependent heuristic dynamic programming(ADHDP)is studied.To overcome the influence of saturated nonlinearity,non-quadratic function is introduced when the system controller is designed.The optimal control strategy is learned by the action network and the critic network in the ADHDP structure.The event-triggered condition determines the weight update time of the network.Finally,the results show that the MAS with input constraints is stable,the state of these agents tends to be consistent,and network resource is also saved.(2)For the optimal consistency control problem of MAS with time-varying input delays,event-triggered HDP is studied.A discretization model is used to convert the original system into a system without time-varying input delay,and it is proved that the system performance index functions before and after the conversion are equivalent.The event-triggered condition is designed to ensure that the model network,action network,and critic network in the HDP structure are only updated when the event is triggered.Finally,it is verified by simulation that this method can guarantee the state of these agents is ultimately uniform.Compared with the time-driven method,it can be seen that the method designed in this paper not only saves network resources,but also ensures better performance of MAS with variable input delays.(3)For the optimal consistency control problem of MAS with input disturbance,event-triggered single network adaptive dynamic programming is designed.When designing the controller,the coupling gain is multiplied by the analytical solution of the cost function in the system to construct a control strategy against the disturbance.The input disturbance term is replaced by a neural network model,which regulates and restricts each other with the action network.The optimal control strategy can minimize the cost function under the premise of maximum input disturbance.The critic-action-disturbance network shares the same weight updating rule derived from the critic module.The network is updated at the moment of event-triggered,which reasonably avoids unnecessary calculation in network learning.The final simulation verifies that the proposed method can come up to the expected performance of the system.In addition,the waste of information resources in the process of communication is decreased.