Event‐Triggered Off‐Policy Learning for Data‐Efficient Control of Complex Unmodeled Systems

ABSTRACT This paper addresses the challenges of low data efficiency and diminished control accuracy in complex systems with completely unknown dynamics. Conventional off‐policy adaptive dynamic programming (ADP) methods, which rely on random data sampling, often lead to the accumulation of irrelevant data, straining computational resources, and hindering learning performance. To overcome these limitations, we propose a novel framework that integrates an event‐triggered mechanism into the off‐policy ADP algorithm. This innovation replaces passive, periodic sampling with an active, intelligent paradigm that collects data only when a prescribed triggering condition is met, thereby ensuring the high informational value of each sample. We provide rigorous theoretical guarantees, including the uniform ultimate boundedness and near‐optimality of the closed‐loop system and exclude Zeno behavior. The efficacy of the proposed approach is corroborated via extensive simulations on an F‐16 aircraft model and other nonlinear systems, demonstrating significant improvements in data efficiency and control performance compared to conventional methods.