EVOLUTIONARY OPTIMIZATION OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURES IN INTELLIGENT DATA ANALYSIS TASKS

Abstract

This paper presents an approach to the automated optimization of convolutional neural network architectures based on evolutionary computation. The main focus is placed on the application of genetic algorithms as one of the most effective classes of evolutionary algorithms capable of performing global search for optimal solutions in complex high-dimensional parameter spaces. The proposed approach is aimed at optimizing the structural parameters of a convolutional neural network, including the number of convolutional and fully connected layers, the number of neurons in each layer, as well as the number and size of convolutional filters, which directly affect the generalization properties of the model. The paper analyzes the fundamental principles of convolutional neural networks and their application to image classifi- cation tasks. The basic components of convolutional neural network architectures are examined, including convolutional layers, subsampling layers, and fully connected layers, and the influence of their parameters on classification accuracy is investigated. Based on this analysis, a baseline convolutional neural network with an architecture manually selected by the developer was implemented, and an experimental evaluation of its performance on both test and real-world datasets was conducted. In addition, a proprietary genetic algorithm was developed on the basis of the fundamental principles of genetic algo- rithms to optimize the architecture of the convolutional neural network. Classification accuracy was employed as the fitness function, allowing the evolutionary search process to be directly guided toward improving the model’s performance. The developed genetic algorithm performs an automated search for an optimal neural network architecture through evolutionary modification of its structural parameters. During the evolutionary process, a gradual improvement of network characteristics is achieved by selecting the most effective configurations and generating new architectures through crossover and mutation operations. Experimental results demonstrate that within ten generations of evolution, it is possible to obtain a convolutional neural network architecture that significantly outperforms the initial manually designed model.