Distributed Optimal Tracking Control of Discrete-Time Multiagent Systems via Event-Triggered Reinforcement Learning

In this paper, an event-triggered optimal tracking control of discrete-time multi-agent systems is addressed by using reinforcement learning. In contrast to traditional reinforcement learning-based methods for optimal coordination and control of multi-agent systems with a time-triggered control mechanism, an event-triggered mechanism is proposed to update the controller only when the designed events are triggered, which reduces the computational burden and transmission load. The stability analysis of the closed-loop multi-agent systems with event-triggered controller is described. Further, to implement the proposed scheme, an actor-critic neural network learning structure is proposed to approximate performance indices and to on-line learn the event-triggered optimal control. During the training process, event-triggered weight tuning law has been designed, wherein the weight parameters of the actor neural networks are adjusted only during triggering instances compared with traditional methods with fixed updating periods. Further, a convergence analysis of the actor-critic neural network is provided via Lyapunov method. Finally, two simulation examples show the effectiveness and performance of the obtained event-triggered reinforcement learning controller.

I. Introduction

Distributed coordination control of multi-agent systems (MASs) has generated a lot of research interest due to its extensive applications to the fields of power systems [1], [2], distributed sensors networks [3], multi-vehicle system [4], multiarea power systems [5], and many others. The objective of coordination control of MASs is to design a distributed controller to achieve an agreement or synchronization through local information interaction among agents. Such control problems of MASs have attracted much attention in a variety of perspectives, for example in synchronization control [6], [7], consensus control [8]–[11], tracking control [12], anti-synchronization control [13], and distributed optimization [14].