Development of a generalized neural network

The interest for neural networks has grown concomitantly with the increased awareness of the ubiquity of non-linear systems. The main focus on improvements in this field has been on the development of different algorithms that either speed up the convergence rate and/or avoid entrapment in local minima. In this work, a different approach is utilized where the existence of local minima is regarded as an exploitable advantage since they can be considered as corresponding to different descriptions of the information content. This study focuses on a method to combine these different descriptions, obtained from several optimized neural networks, into a generalized neural network. The development of generalized neural networks is illustrated using two real-life data sets. The results show that the generalized neural networks improves the estimated Mean Squared Error (MSE) by at least 23%. Furthermore, the generalized neural network does not overfit the calibration set, as the Mean Squared Error of Calibration (MSEC) set is in close agreement with the MSE of the independent test set.