Identification of MIMO systems using MLP networks: Comparison between SVR and random initialisation

Abstract

Neural network (NN) modelling approach is often used for non-linear system identification. Building a NN for some identification problem starts by choosing its structure and initial weights. There is no exact method to determine the optimal initialisation for a NN, but some authors have used support vector regression (SVR) to initialise a RBFNN which could be considered as a systematic way. This paper presents a SVR initialisation method for Multi-Layer Perceptron (MLP) NN. The proposed method is based on the analogy between NN and SVR to determine the necessary number of hidden neurons and the initial weights for a given modelling precision. Simulation results for multi-input multi-output (MIMO) system show the feasibility and accuracy of the proposed method.