USING MODULAR NEURAL NETWORKS AND MACHINE LEARNING WITH REINFORCEMENT LEARNING TO SOLVE CLASSIFICATION PROBLEMS

Context. The solution of the classification problem (including graphical data) based on the use of modular neural networks and modified machine learning methods with reinforcement for the synthesis of neuromodels that are characterized by a high level of accuracy is considered. The object of research is the process of synthesizing modular neural networks based on machine learning methods with reinforcement. Objective is to develop a method for synthesizing modular neural networks based on machine learning methods with reinforcement, for constructing high-precision neuromodels for solving classification problems. Method. A method for synthesizing modular neural networks based on a reinforcement machine learning approach is proposed. At the beginning, after initializing a system of modular neural networks built on the bottom-up principle, input data is provided - a training set of data from the sample and a hyperparameter to select the size of each module. The result of this method is a trained system of modular neural networks. The process starts with a single supergroup that contains all the categories of the data set. Then the network size is selected. The output matrix is softmax, similar to the trained network. After that, the average probability of softmax is used as a similarity indicator for group categories. If new child supergroups are formed, the module learns to classify between new supergroups. The training cycle of modular neural network modules is repeated until the training modules of all supergroups are completed. This method allows you to improve the accuracy of the resulting model. Results. The developed method is implemented and investigated on the example of neuromodel synthesis based on a modular neural network for image classification, which can later be used as a model for technical diagnostics. Using the developed method significantly reduces the resource intensity of setting up hyperparameters. conclusions. The conducted experiments confirmed the operability of the proposed method of neuromodel synthesis for image classification and allow us to recommend it for use in practice in the synthesis of modular neural networks as a basis for classification models for further automation of tasks of technical diagnostics and image recognition using big data. Prospects for further research may lie in using the parallel capacities of GPU-based computing systems to organize directly modular neural networks based on them.