Encoding a priori information in neural networks to improve its modeling performance under non-stationary environment

A possibility of utilizing priori information to improve modeling performance of neural networks under non-stationary environment is examined in this paper. Because the Green function under critical state expresses varying behaviors of the systems, a time series model can be used to depict the time-varying behaviors. The extended Kalman filter (EKF) is often used to train neural networks under non-stationary environment. By a little transformation of time series model, a parameter transfer matrix /spl phi/(k) for the EKF estimation (TSEKF) is derived to deal with the varying characteristics of the systems. The performance of TSEKF algorithm is compared with the existing EKF algorithm, which does not consider any priori information and only assumes that varying characteristics of the systems is random walk. Simulation result indicates that the modeling performance of neural networks under non-stationary environment is remarkably improved for encoding priori information.