Layer-Wise Dynamic Event-Triggered Neural Network Control For Discrete-Time Nonlinear Systems

Recent efforts in the control community focused on developing methods to guarantee closed-loop stability of systems controlled by multilayer perceptron (MLP)-based policies. However, little attention has been paid to the computational demand of such controllers due to frequent network evaluations. In this paper, we address this challenge by proposing a dynamic Event-Triggering Mechanism (ETM) to reduce the computational burden. Specifically, we focus on the stabilization of discrete-time Lur’e systems with input saturation. The proposed strategy reduces the evaluation frequency of the layers of the neural controller while preserving stability guarantees. The ETM is constructed using Linear Matrix Inequality (LMI)-based conditions, which leverage the known properties of activation functions and employ Finsler’s lemma to reduce conservativeness. Numerical results demonstrate the effectiveness of the proposed method, achieving significant computational savings compared to state-of-the-art solutions.