Relaxation Subgradient Algorithms with Machine Learning Procedures

In the modern digital economy, optimal decision support systems, as well machine learning are becoming an integral part of production processes. Artificial neural network training other engineering problems generate such high dimension that difficult to solve with traditional gradient or conjugate methods. Relaxation subgradient minimization methods (RSMMs) construct a descent direction forms obtuse angle all subgradients current minimum neighborhood, which reduces problem solving systems inequalities. Having formalized model and taking into account specific features sets, we reduced system inequalities approximation obtained efficient rapidly converging iterative algorithm for finding descent, conceptually similar least squares method. The new is theoretically substantiated, estimate its convergence rate depending on parameters set. On this basis, have developed substantiated RSMM, has properties method quadratic functions. We practically realizable version uses rough one-dimensional search. A computational experiment complex functions in space confirms effectiveness proposed algorithm. models, where it required remove insignificant variables neurons using Tibshirani LASSO, our outperforms known