Robust adaptive neurofilters with or without online weight adjustment

This paper proposes the use of risk-sensitive criteria for synthesizing a neurofilter that is adaptive to adaptation-worthy environmental parameters and robust to adaptation-unworthy ones. Two types of robust adaptive neurofilter are presented, one requires online weight adjustment and the other does not. A robust adaptive neurofilter without online weight adjustment is a time lagged recurrent network (TLRN) synthesized from realizations of the signal and measurement processes at typical values of the adaptation-worthy environmental parameters in a priori off-line training with respect to a risk-sensitive training criterion. A robust adaptive neurofilter with online weight adjustment consists of a signal estimator, a measurement predictor and a weight transformer, the former two being TLRNs and the latter a feedforward ANN. The nonlinear and linear weights of the signal estimator and measurement predictor are used as their long and short-term memories respectively. The linear weights of the measurement predictor are adjusted online by risk-sensitive or H^∞ algorithms, and then transformed into the linear weights of the signal estimator by the weight transformer