V.V. Ayuyev

The paper describes an original model - DEC RBF - for the radial-based neural networks training. The model-s work is based on density-based erosion clustering algorithm. The minimal spanning tree algorithm for cluster size optimization processes initial data domains. Missing data could be imputed independently over each cluster. BF centers are set into the corresponding cluster centers. Three different domain databases were used as a testing area. An open-access experimental data were differ both by size and attribute number. Comparative analysis of the proposed model, traditional neural network based architecture, and previously described model was fulfilled. The results showed a slight lack of DEC RBF model-s accuracy in comparison with the best-known solution (which was our previous model), that is due to the much lesser number of BF. In the case of similar amount of BF in a neural network architecture, DEC RBF showed 1,3-2 times better accuracy along with 2-3,5 better performance. The optimal value of internal parameter was found because of analysis for exogenous and endogenous factors influence on model-s quality rates. One of the key features of the proposed method was fully repeatable result for network train-ing.