Orthogonal iterative learning least squares for neural identification of nonlinear time-varying systems

By utilizing QR decomposition technique, an orthogonal iterative learning least squares algorithm is proposed for time-varying high-order neural network training, which is applied for the identification of time-varying nonlinear systems over a finite time interval. With the help of two-dimensional Givens transformation, both on-line and off-line identification procedures are presented for weights update in an iterative manner. Numerical results are given which verify that time-varying weights converges as iteration number increasing, and the neural network output can follow the practical output data.