Core Concepts
This work presents a generalized framework for recursive least squares (GF-RLS) that encompasses many extensions of RLS as special cases. It provides stability guarantees for fixed parameter estimation and robustness guarantees for time-varying parameter estimation in the presence of measurement and regressor noise.
Abstract
The key highlights and insights of this content are:
Derivation of Generalized Forgetting Recursive Least Squares (GF-RLS):
GF-RLS is a discrete-time generalization of the continuous-time RLS framework developed in prior work.
It encompasses many existing extensions of RLS as special cases, allowing a unified analysis.
The GF-RLS cost function is composed of a loss term, a forgetting term, and a regularization term.
Stability Analysis for Fixed Parameter Estimation:
Sufficient conditions are provided for Lyapunov stability, uniform Lyapunov stability, global asymptotic stability, and global uniform exponential stability of the parameter estimation error in GF-RLS.
The stability conditions relate to the properties of the forgetting matrix, covariance matrix, and regressor sequence.
These stability results generalize and extend previous analyses of RLS extensions.
Robustness Guarantees for Time-Varying Parameter Estimation:
Robustness guarantees are derived for GF-RLS in the presence of time-varying parameters, measurement noise, and regressor noise.
Sufficient conditions are provided for the global uniform ultimate boundedness of the parameter estimation error.
A specialization of this result gives a bound on the asymptotic bias of least squares estimators in the errors-in-variables problem.
Survey of RLS Extensions as Special Cases of GF-RLS:
Ten different extensions of RLS, including exponential forgetting, variable-rate forgetting, and directional forgetting, are shown to be special cases of GF-RLS.
This unifies the analysis of these RLS extensions and demonstrates the generality of the GF-RLS framework.
Overall, this work provides a comprehensive theoretical foundation for the analysis of various RLS extensions, with guarantees on stability, robustness, and connections to the errors-in-variables problem.