Liubov S. Lomakina, Igor D. Chernobaev, Liubov A. Beliaeva

Nizhny Novgorod state technical university n.a. R. E. Alekseev

The paper considers the machine learning task of classifying the states of objects with a complex structure. The state of an object is understood as a certain category that characterizes the properties of an object at a given time, and at the same time is described by a set of features. The complexity of the object structure is expressed by the features that require preliminary processing and have hierarchical or sequential relationships, or unstructured elements. The data to process may contain inaccuracies, and errors. The neural networks as a tool are often applied to the classification tasks. Fuzzy logic systems are applicable to the classification tasks within the context of data fuzziness. Both tools have the properties of a universal approximator, however the neural network results are hard to interpret, at the same time the application of a fuzzy logical system requires the preliminary construction of a fuzzy rules. The paper considers an approach to combine fuzzy logic system and neural network with the fuzzy neuron activation function. The model of the fuzzy function is described by the fuzzy sets, membership functions and fuzzy activation rules. The input of a fuzzy function goes through the transformations stages, including fuzzification, fuzzy inference using the fuzzy logic system’s rules and defuzzification. The approach makes it possible to form a fuzzy activation function with a set of changeable parameters during the neural network model training. An improved normalization of the input of the fuzzy activation function is proposed in order to ensure the feasibility of the information signal requirements in the fuzzy function. Neuro-fuzzy classification systems are analyzed in the tasks classification of objects of various nature: technical objects, biomedical objects textual objects. A comparative assessment of the accuracy of solutions of neuro-fuzzy classification systems relative to similar ANN-based systems showed an increase from 2 to 9% in a number of experiments.

classification of objects states, fuzzy logic system, neural networks, neuro-fuzzy classification