Robust adaptive identification of dynamic systems by neural networks

This paper is concerned with the use of neural networks for robust and adaptive identification of dynamic systems. Two types of neural identifiers, that are robust to adaptation-unworthy environmental parameters and adaptive to adaptation-worthy ones, are discussed, one requiring online weight adjustment and the other not. Risk-sensitive criteria are proposed for both training neural identifiers off-line and adjusting their weights online. These criteria induce robust performance by emphasising large errors in an exponential manner. A robust adaptive neural identifier without online weight adjustment is a time lagged recurrent network (TLRN) synthesized from input/output data of the dynamic system at with respect to a risk-sensitive criterion. The neural identifier´s ability to adapt to the adaptation-worthy environmental parameters is a manifestation of the ability of the TLRN to estimate these parameters internally. A robust adaptive neural identifier with online weight adjustment is a neural network with long- and short-term memories. The long-term memory, which consists of the nonlinear weights of the neural network, is determined in an a priori off-line training in such a way that it is independent of the adaptation-worthy variables. The short-term memory, which consists of the linear weights of the neural network, is adjusted online to adapt to the adaptation-worthy variables. The criteria used for both the off-line determination of the long-term memory and the online adjustment of the short-term memory are risk-sensitive to induce the neural identifier´s robust performance