Delay-Enabled Adaptive Learning of Time-Varying Parameters

Abstract

This paper proposes a novel adaptive learning method for unknown time-varying parameters by deliberately incorporating a delay into the parameter update. Specifically, a conventional adaptive estimator is augmented with only one additional dynamic state, which estimates the parameter variation rate from the difference between current and previous parameter estimates. The method fundamentally changes the influence of time variation on estimation performance: the resulting estimation error is governed by the second derivative of the unknown parameter rather than its first derivative. Under persistent excitation, a small-gain analysis proves exponential stability of the complete estimation-error system and establishes explicit bounds for the parameter, state-prediction, and parameter-rate errors. In particular, accurate learning of linearly time-varying parameters and their unknown rates can be guaranteed with errors converging exponentially fast to zero. Numerical comparisons with conventional adaptive estimation and a recent timevarying parameter estimator demonstrate superior learning performance in the considered examples. The proposed delay-enabled adaptive estimator requires neither state derivatives nor any assumed model of parameter variation, and its implementation consists of one delayed difference and a first-order filter. These results thus provide a direct extension of adaptive learning from constant to time-varying parameters without sacrificing the simplicity of the conventional adaptive estimation mechanism.