Genetic Optimization of Convolutional Neural Network Architectures Using Topological Crossover

This work explores the possibility of automatically searching for the structure and hyperparameters of shallow convolutional neural networks using genetic algorithms, allowing for higher accuracy than that of manually compiled models. The purpose of the article is to develop a genome structure for creating a neural network, describe the method of operation of a genetic algorithm that allows individuals to train, describe and analyze methods of mutation and crossing genomes to obtain a population with a high level of diversity. The analysis of the learning outcomes of the genetic algorithm, the complexity of its application to generate effective neural network structures, and describes its own method for topologically crossing convolutional neural network structures. The final genetic algorithm in combination with the studied method of crossing neural networks allowed to improve the convergence of the population, and the resulting best model, which has a non-trivial internal structure, was able to achieve an accuracy of 92.91% in the test sample, which is on average 2% higher than that of manually created models.