Dynamic multivariate B-spline neural network design using orthogonal least squares algorithm for non-linear system identification

This paper investigates the design of a multivariate B-spline neural network using the orthogonal least squares algorithm for non-linear system identification. The B-spline neural network is a type of basis function neural network which has been developed from the function approximation approach based on B-spline functions. Usually, this kind of neural network is trained using the gradient-based algorithm. In order to overcome the problems regarding the stability of the training procedure, a learning procedure based on the orthogonal least squares algorithm is suggested in this paper. This approach allows for the setting of the B-spline functions knots and also of the neural network´s weights. An experimental study was conducted in order to identify a neural model for the Three-Tank System laboratory set-up using the multivariate B-spline neural network designed with the suggested procedure.