Estimation of weights to combine trained neural networks using linear estimation techniques

Input feature selection and appropriate weight estimation in combining trained neural network are one of the important key factors in neural stacked neural network based models. This study has investigated these problems through the information theory and linear estimation weight techniques. These techniques have implemented to model the non linear separation process model of a novel design compact separator. The separation efficiency of the compact separator used in this study was found to be dependent non -linearly on many input factors such as gas volume fraction, inlet mixture velocity, liquid and gas superficial velocity, inlet pressure, and Loss coefficient etc. The input parameters from all the candidate inputs were selected based on their Mutual Information with the separation efficiency. It is demonstrated that mutual information is better statistical method for input feature selection to train a neural network. Based on the mutual information a set of inputs were selected and several single trained neural networks having different architecture in terms of hidden neurons and training functions were combined together to improve the prediction accuracy. Three linear methods i.e. equal weigh ts; linear regression and principle component regression were used to combine the trained neural network. The performance of the combined neural network aggregated through principle component regression was found to better than neural network combined with equal weight and linear regression.