Attention-Enabled Memory for Concurrent Learning Adaptive Control

Abstract

Transient tracking error dynamics are inevitable in any practical closed-loop control system. While numerous works are devoted to improving these dynamics, in this letter, we focus on taking advantage of it first, in the context of adaptive control. We propose a memory architecture that can make use of stored significant data about the transients of previously experienced anomalies to aid in obtaining a resilient system against uncertainties. The proposed architecture consists of 1) a memory containing data about a variety of uncertainties, 2) a short-term memory that aids in handling new uncertainties, and 3) an attention-based reading mechanism that enables the controller to retrieve only relevant data from the memory. The effectiveness of the architecture is validated through numerical simulations, and a rigorous Lyapunov stability analysis is provided.