Event-triggered ADP for asymptotic output regulation of unknown linear systems

Abstract

The deployment of learning-based controllers in modern networked cyber-physical systems is constrained by bandwidth limitations, partial state observability, and parametric uncertainties. Traditional adaptive dynamic programming (ADP) relies on full-state feedback and periodic sampling, inducing network congestion and steady-state tracking errors. This paper develops a fully data-driven dynamic event-triggered output regulation scheme operating on measurable input-output streams. By integrating a projection-safeguarded value iteration algorithm with an internally evolving dynamic threshold variable, we guarantee asymptotic convergence of the sampling error without requiring an admissible initial stabilizing policy or explicit system identification. Simulations on a third-order uncertain plant demonstrate a 54.0% reduction in control transmissions and an 8.2% suboptimality bound relative to the model-based LQR baseline, with reconstruction error consistently decaying below 10^-3. The framework provides a computationally efficient, theoretically rigorous architecture for resource-constrained cyber-physical networks.