Comparison of Genetic and Bayesian Neural Architecture Searches for Anomorphosis Approximation Problem

This paper investigates optimal configuration of three neural network architectures using two methods. The neural network solves the problem of approximating the dynamic anomorphosis method using experimental data. The method uses the original visual image and transforms it based on the Euclidean metrics into another image according to a given criterion. To learn the neural networks, a dataset of several tens of thousands of pairs consisting of visual image and non-regular transformation raster, which is the algorithm result from the previous works, was collected. A new approach to expanding the learning sample through aligned augmentation of the input and output data is proposed. Differences and similarities between the modified genetic algorithm and the Bayesian search are investigated. Both methods find optimal configurations at approximately the same speed, while the found configurations of the hybrid genetic algorithm being optimal by the Pareto principle appear to be more diverse. Selecting architecture of the neural network results in a model with determination coefficient (R2) between the reconstructed and original images equal to 0.81.