Selective Ensemble Random Neural Networks Based on Adaptive Selection Scope of Input Weights and Biases for Building Soft Measuring Model

摘要

Random neural networks (RNNs) prediction model is built with a specific randomized algorithm by employing a single hidden layer structure. Duo to input weights and biases are randomly assigned and output weights are analytically calculated, it is widely used in different applications. Most of RNNs-based soft measuring models assign the random parameter scope to default range [- 1, 1]. However, this cannot ensure the universal approximation capability of the resulting model. In this paper, selective ensemble (SEN)-RNN algorithm based on adaptive selection scope of input weights and biases is proposed to construct soft measuring model. Bootstrap and genetic algorithm optimization toolbox are used to construct a set of SEN-RNN models with different random parameter scope. The final soft measuring model is adaptive selected in terms of the best generation performance among these SEN models. Simulation results based on housing benchmark dataset of UCI and dioxin concentration dataset of municipal solid waste incineration validate the proposed approach.