Improving Nonlinear Process Modelling Through Selective Combination of Multiple Neural Networks using Combined Correlation Coefficient Analysis

Abstract

This paper proposed a selective combination method based on combined correlation coefficient analysis to increase the robustness of the single neural network. The main objective of the proposed approach is to improve the generalisation capability of the neural network models by combining networks that are less correlated. The assumption that we made is that combining networks that are highly correlated might not improve the final prediction performance due to the fact that these networks present the same contribution to the final prediction. This might even deteriorate the robustness of the combined network. The result shows that combination multiple neural networks using the proposed approach improved the performance of the two nonlinear process modelling case studies in which there is a significant reduction of validation sum square error (SSE) of the networks was obtained. Key words: Multiple neural networks, selective combination neural networks, correlation coefficient, nonlinear process modelling