Experience Replay Enhances Excitation Condition of Neural-Network Adaptive Control Learning

Neural-network adaptive control offers a promising avenue for managing dynamic uncertainties with unknown structures. Nevertheless, traditional neural-network adaptive control methodologies are plagued by challenges such as persistent excitation and learning singularities. This study introduces an enhanced neural-network adaptive control methodology that optimizes the use of empirical data and ensures a more even spatial distribution of samples. A neural-network adaptive controller is developed, incorporating an experience replay strategy that refines the adaptive tuning of network weights. The adaptation is fueled by the current tracking and historical model prediction errors, markedly improving the excitation condition. Furthermore, a data- selection strategy that prioritizes regions with sparse samples is introduced, effectively mitigating the issue of learning singularities. Additionally, an adaptive technique to circumvent control saturation is employed. The stability of the control system and the convergent behavior of the network weights are substantiated through theoretical analysis. Simulation outcomes corroborate the superiority of the proposed algorithm in terms of parameter convergence, learning efficacy, and control precision.