Neural networks initialization

The paper investigates the problem of neural network initialization for the identification of linear time invariant dynamical systems. The multilayer feedforward network with linear neurons associated with multiple delay lines is used to perform identification of arx models. Special attention is devoted to the initialization of weights using a priori knowledge on the model structure and parameters, and to robustness performance of neural models. Simulation results enhance the weakness of random initial weights on learning and highlight the performance of optimized initial weights on robustness. Some indications are given for the implementation of the aforementioned initialization