Model identification of non-linear dynamical system by recurrent neural networks

Methods for model identification are crucial in many fields, such as adaptive signal processing, pattern classification, adaptive control. In these cases, the central problem is to seek an artifical model, capable of producing numerical outputs that match the corresponding values measured from a physical system. In this paper we propose an approach to identification of nonlinear dynamic system which usesecurrent neural networks. In particular our approach globally identifies the observed system without any need of separating the contriburion of linear and nonlinear parts, which requires an a priori knowledge of the internal system structure. The proposed approach is used in several problems, such as system identification, time series prediction and control. Experimental results both in classical synthetic examples and practical problems prove the efficiency of the method.