Reinforcement-learning-based decentralized event-triggered control of partially unknown nonlinear interconnected systems with state constraints

Abstract

In many applications with great potential, safety is critical as it needs to meet strict safety specifications within physical constraints. This paper studies the decentralized event-triggered control problem of a class of partially unknown nonlinear interconnected systems with state constraints under the reinforcement learning approach. First, by introducing a control barrier function into the performance function of each auxiliary subsystem with state constraints, the system state can be operated within a user-defined safe set. And then, the original control problem can be translated equivalently into finding or searching optimal event-triggered control policies that combine to form the desired decentralized controller, resulting in significant savings in communication resources. Compared with the traditional actor-critic network structure approach, the proposed identifier-critic network structure can loosen the constraints on the system dynamics and eliminate the errors arising from approximating the actor network. Updating the weight vectors in the critic network by gradient descent and concurrent learning techniques removes the need for the traditional persistence of excitation conditions. Furthermore, it is rigorously proved that all the signals of the interconnected nonlinear system are bound according to the Lyapunov stability theory. Last, the effectiveness of the proposed control scheme is verified by simulation examples.