ESO-based Data-driven Learning Control For Multi-agent Networks

In this thesis,a series of ESO-based data-driven learning control methods are proposed for the problem of multi-agent consensus under different network topologies with repeated operation by combining with an improved dynamic linearization model.The effectiveness of the proposed methods is verified through the strict mathematical analysis and studies.The main innovations of this thesis are summarized as follows:First,a data model is designed to establish the dynamic relationships of the agents from both iterative domain and network topology perspectives for single sequence topological multi-agent systems,which includes a linear parameter term and a nonlinear residual term.The dynamic relationship is not constrained by the initial state of the agent and does not involve vector computation,which reduces the computational burden of the controller afterwards.Based on this,an optimal learning control law and an adaptive learning control law are designed for control tasks,i.e.,fixed formation goal and varying formation goal,respectively.In addition,the learning capability is improved by considering the control information of the parent agent in spatial dimension,and the ESO is used to estimate the unknown nonlinear term in the controller,which reduces the complexity of the linear parameter and improves the integrated control performance.Second,based on the spatial dynamics proposed for the single sequence network,the dynamic relationship of each agent is established for strongly connected multi-agent systems by considering the strongly connected topology and combining the nonlinear uncertainties and external disturbances among the agents.On this basis,an ESO based data-driven learning control method is proposed.The operation of the control process established in this way does not depend on the initial state information of the agents,which improves the ability of the system to cope with non-repetitive uncertainty.Moreover,the controller does not involve vector computation,so the controller will take up less computing resources.The method is also able to learn additional control experience from neighbor agents,and the introduction of ESO enhances the anti-disturbance of the control method and improves the control performance.Third,an ESO based data-driven adaptive iterative learning bipartite consensus protocol is proposed for cooperation-antagonism multi-agent systems.The dynamic relationship between different batches is established at each operation point for each individual agent,which transforms the plant into an affine structure with a linear parameter term and a nonlinear residual term.The resulting linear data model is not constrained by the initial state of the agents and does not involve vector computation,which means that the pressure on the computing unit will be reduced.On this basis,an adaptive iterative learning protocol is proposed by combining the bipartite consistency network topology.The control gain is dynamic,which can be adaptively adjusted by input and output data.In addition,the control performance is improved by estimating the nonlinear residual term through ESO.