Multiple-model approach to finite memory adaptive filtering

The multiple-model technique is proposed for the purpose of finite memory adaptive filtering of nonstationary signals. Its most important feature is the parallel structure of computation: not one but several identification algorithms characterized by different memory-controlling parameters are run in parallel and combined appropriately. The results substantially improve the robustness of the adaptive scheme to the experimenter's choice of design parameters such as forgetting factors, adaptation gains, and model orders. The author suggests a technique which allows for a rational decision to be made when several competitive adaptive filters work simultaneously. The results obtained can also be used for the purpose of model order determination, and their close correspondence to Rissanen's predictive least squares principle and Akaike's concept of model likelihoods is noted. >