Development of an Algorithm for Multicriteria Optimization of Deep Learning Neural Networks

Nowadays, machine learning methods are actively used to process big data. A promising direction is neural networks, in which structure optimization occurs on the principles of self-configuration. Genetic algorithms are applied to solve this nontrivial problem. Most multicriteria evolutionary algorithms use a procedure known as non-dominant sorting to rank decisions. However, the efficiency of procedures for adding points and updating rank values in non-dominated sorting (incremental non-dominated sorting) remains low. In this regard, this research improves the performance of these algorithms, including the condition of an asynchronous calculation of the fitness of individuals. The relevance of the research is determined by the fact that although many scholars and specialists have studied the self-tuning of neural networks, they have not yet proposed a comprehensive solution to this problem. In particular, algorithms for efficient non-dominated sorting under conditions of incremental and asynchronous updates when using evolutionary methods of multicriteria optimization have not been fully developed to date. To achieve this goal, a hybrid co-evolutionary algorithm was developed that significantly outperforms all algorithms included in it, including error-back propagation and genetic algorithms that operate separately. The novelty of the obtained results lies in the fact that the developed algorithms have minimal asymptotic complexity. The practical value of the developed algorithms is associated with the fact that they make it possible to solve applied problems of increased complexity in a practically acceptable time. © Authors retain all copyrights.