Reinforcement Learning for Robust Dynamic Event-Driven Constrained Control | JoVE Visualize

Area of Science:

Control Systems Engineering
Nonlinear Dynamics
Artificial Intelligence in Control

Background:

Nonlinear systems often face challenges with unmatched perturbations and unknown, asymmetric/symmetric input constraints.
Existing control methods may struggle with computational load and stability under such complex conditions.
Event-driven control (EDC) offers a promising approach to reduce computational burden by triggering control updates only when necessary.

Purpose of the Study:

To develop a robust dynamic event-driven control (EDC) strategy for nonlinear systems with unmatched perturbations and unknown input constraints.
To reduce the computational load of control systems through a novel dynamic event-triggering mechanism.
To ensure the stability and robustness of the closed-loop system under challenging conditions.

Main Methods:

Construction of a novel nonquadratic cost function for a constrained auxiliary system to handle symmetric or asymmetric input constraints.
Proposal of a dynamic event-triggering mechanism utilizing time-based variables and system states.
Development and solution of the event-driven Hamilton-Jacobi-Bellman equation using a critic neural network (CNN) within a reinforcement learning framework, incorporating experience replay to relax excitation conditions.
Lyapunov's approach for stability analysis of the closed-loop auxiliary system and weight estimation error.

Main Results:

Demonstrated that robust dynamic EDC for the original system can be achieved by solving the event-driven optimal control problem of the auxiliary system.
Validated the uniform ultimate boundedness stability of the closed-loop auxiliary system and the weight estimation error.
Successfully applied the method to a nonlinear plant and a pendulum system, confirming theoretical claims.

Conclusions:

The proposed robust dynamic EDC strategy effectively handles nonlinear systems with unmatched perturbations and complex input constraints.
The dynamic event-triggering mechanism significantly reduces computational load while maintaining system stability.
The critic neural network approach within reinforcement learning provides a viable solution for solving the complex control problem.