Two-Tier Filter-based Experience Replay for Neuro-Adaptive Control with Inherent Robustness

Abstract

Neural network (NN) - based adaptive control is popularly used for controlling nonlinear dynamical systems with unstructured uncertainty. NN structures are embedded in the feedback loop with a real-time tuning mechanism to ensure desired tracking performance. Due to the presence of approximation error, neuro-adaptive controllers are affected by the phenomenon of parameter drift in the absence of persistence of excitation (PE) condition, and thereby require robust modifications like projection, σ-modification etc. to guarantee boundedness of the estimation error. The PE condition is stringent in nature since it lacks practical feasibility in most real world applications. Unlike conventional neuro-adaptive controllers, the proposed work relaxes the need for PE condition, while imposing a milder condition, called initial excitation (IE) for ensuring robust tracking. The algorithm relies on a dual-layer filter architecture to facilitate the NN weight tuning mechanism, while inserting an experience replay block in the update law based on captured information from IE condition. The proposed algorithm has a feature of inherent robustness, i.e., it does not require any robust modification like the traditional neuro-adaptive controllers. The IE-based experience replay suffices to prevent parameter drift by ensuring uniform ultimate boundedness (UUB) of the closed-loop concatenated error (tracking error+ NN weight tuning error) dynamics. A Barrier Lyapunov Function (BLF) is introduced in the adaptive law, which facilitates compact set pre-conditioning for the applicability of Universal Approximation property of the NN structure. Moreover, the ultimate-bound can be arbitrarily reduced by selecting the design parameters appropriately. Simulation results on wing-rock dynamics validate the efficacy of the proposed method as compared to status quo.