This paper introduces a novel framework for Online Adaptive Dictionary Refinement (OADR) that enables real-time system identification through sparse representation learning with dynamic dictionary adaptation. The proposed method integrates recursive least squares with online dictionary learning to estimate system parameters and refine the underlying representation basis. A coherence-based atom selection criterion is introduced to ensure dictionary diversity and maintain computational efficiency suitable for real-time deployment.
Key findings
Proposes OADR framework combining recursive parameter estimation with online dictionary adaptation.
Introduces coherence-based atom selection mechanism for dictionary diversity.
Theoretical analysis provides convergence guarantees and error bounds.
OADR outperforms conventional adaptive filtering methods in tracking time-varying systems.
Limitations & open questions
Assumes parameter vector can be represented sparsely in a dictionary.
Fixed model structures may be insufficient for capturing highly time-varying dynamics.