Structure relaxation method for self-organizing neural networks

Self-organizing neural networks achieve more predictable and accurate results then the classic ones with the static architecture. Neurons and connections of such neural networks are dynamically built during the learning process. Self-organizing neural networks based on the group method of data handling (GMDH) have proven to be one of the most efficient approaches to solving the problems of pattern recognition with the statistical learning data. In this article we propose a new method for searching deeper interrelations of the inputs and the output of the system under the study of such a neural network. The method allows eliminating links to the inputs that are no longer useful at the later steps of the neural network construction, thus allowing to simplify the neural network structure and increase prediction accuracy. Hence the method is called the structure relaxation method. For complex problems the method helps to find deeper system inputs interrelations, increase the prediction accuracy, and, at the same time, decrease the number of the inputs being used. The proposed relaxation method was tested on the real world problems; the results are also presented herein.