Optimal selection of learning parameters for regularised random vector functional-link networks-based soft measuring model

Abstract

Random vector functional-link networks (RVFLNs) with one single hidden layer structure have been used widely for soft measuring model construction. In which, the input weights and biases are produced randomly and the output weights are computed analytically by a Moore-Penrose generalised inverse method. Regularised RVFLN (RRVFLN) can prevent over-fitting problem and reduce complexity of the constructed model by using the ridge regression method. Several learning parameters, such as range of random input weights and bias, number of hidden nodes and regularising factor are data dependent. This paper aims to develop a composite differential evolution (CoDE)-based optimal selection method to address the three learning parameters of RRVFLN. Experiments on some benchmark datasets are carried out to validate the proposed method.