Synthesis of radial basis neural networks with the transformation of a generalized axis to automate diagnostic decisions

The problem of synthesis of radial basis neural networks based on a set of precedents for the automation of decision-making in the diagnosis is solved. The method of radial basis neural network synthesis is proposed. It uses a transformation mapping a sample from the multidimensional feature space into one-dimensional generalized axis to allocate cluster centers and boundaries and, in contrast to the known methods, does not require the user's involvement in clusters number selection, has no uncertainty in the selection of the number of neurons in the first layer and the in the choice of the initial weight values of the network, aims to minimize the size of the network, characterized by an acceptable training time, at the clusters partition formation takes into account the feature informativities, through the use of network optimization procedure allows to obtain non-redundant interpretable contrast models, and allows additional training of previously built models. The software that implements the proposed method is developed. Experiments were conducted to study the method. They confirmed its efficiency and allow to recommend the method to be used in practice. It is found that the proposed method is characterized by a less time-consuming to build a radial basis neural-networks compared to using clear exhaustive search of the cluster analysis and time-consuming in comparison with the mapping samples in clusters. Thus the proposed method provides acceptable accuracy and significantly higher level of abstraction than the known methods. Key words: neural network, radial base network, training, synthesis, diagnostics.