Improving the Classification Accuracy of Multidimensional Overlapping Objects Based on Neuro-Fuzzy Analysis of Generalized Patterns

Abstract

To improve the classification accuracy of multidimensional overlapping objects, a new hybrid neuro-fuzzy FCNN-SOM-FMLP network, combining the fuzzy cell neural network of Kohonen (FCNN-SOM) and the fuzzy multilayer perceptron (FMLP), and the algorithms for its training are proposed. This combination allows for clustering of generalized intersecting patterns (the extensional approach) and training the classification network basing on the identification of integrated pattern characteristics in the isolated clusters (intentional approach). The new FCNN-SOM-FMLP architecture features a high degree of self-organization of neurons, an ability to manage selectively individual neuronal connections (to solve the problem of “dead” neurons), the high flexibility, and the ease of implementation. The experimental results show the temporal efficiency of algorithms of self-organization and training and the improvement of the separating properties of the network in the case of overlapping clusters. Calculated technological and economic generalized values of countries.