Robust adaptive neurofilters with or without online weightadjustment

摘要

This paper proposes the use of risk-sensitive criteria fornsynthesizing a neurofilter that is adaptive to adaptation-worthynenvironmental parameters and robust to adaptation-unworthy ones. Twontypes of robust adaptive neurofilter are presented, one requires onlinenweight adjustment and the other does not. A robust adaptive neurofilternwithout online weight adjustment is a time lagged recurrent networkn(TLRN) synthesized from realizations of the signal and measurementnprocesses at typical values of the adaptation-worthy environmentalnparameters in a priori off-line training with respect to anrisk-sensitive training criterion. A robust adaptive neurofilter withnonline weight adjustment consists of a signal estimator, a measurementnpredictor and a weight transformer, the former two being TLRNs and thenlatter a feedforward ANN. The nonlinear and linear weights of the signalnestimator and measurement predictor are used as their long andnshort-term memories respectively. The linear weights of the measurementnpredictor are adjusted online by risk-sensitive or H^∞nalgorithms, and then transformed into the linear weights of the signalnestimator by the weight transformer