A nonlinear adaptive fuzzy approximator technique with its application to prediction of non-stationary EEG dynamics and estimation of single-sweep evoked potentials.

In this paper, we apply a fast training paradigm to the optimization of fuzzy approximator and a nonlinear adaptive fuzzy approximator (NAFA) is constructed. Using TSK fuzzy rules, the structured knowledge along with numerical information are parameterized and utilized in the NAFA, which can be easily configured as a multi-layer network when its transparency is desired. We propose a fast training paradigm, which is actually a combination of Kalman filtering and LMS adaptation, to optimize the linear and nonlinear parameters of the NAFA separately. The NAFA is characterized by concise representation of structured knowledge, fast learning capability, as well as universal approximation property. The NAFA is applied to forecast the non-stationary EEG time-series and to estimate single-sweep evoked potentials (EPs). The corresponding simulation results are given. It is concluded that the NAFA technique can provide efficient nonlinear separation of single-sweep EPs, which allows for quantitative examination of the cross-trial variability of clinical EPs.