Adaptive IIR filtering: composite prefiltered regressor method

In this paper, we present a new method for adaptive IIR filtering that promises improved performance under general operating conditions, including both colored and white noise, sufficient and insufficient order filter cases. By defining the regressor in adaptive IIR filter update to be a convex combination of the regressors for the Steiglitz-McBride method (SMM) and recursive predicted error method (RPEM), we are able to tradeoff the benefits of each. RPEM minimizes mean square output error (MSOE) directly, and thus has slow convergence rate and may converge to a local minimum because of nonconvexity and multimodality of MSOE surface, SMM converges fast, but may converge to a biased solution or diverge in colored noise environments. Other composite methods (e.g., composite regressor method (CRM)) use a similar approach, but only focus on reducing bias of equation error estimates for sufficient order filters in white noise environments. Conversely, our method, CPRM, can extend applications to general environments, prevent diverging, reduce bias, and increase likelihood of convergence to the global minimum