On the convergence of a training algorithm based on the Boltzmann annealing optimization scheme
V. V. Krasnoproshin, V. V. Matskevich · Digital Library of the Belarusian State University (Belarusian State University) · 2025
The paper deals with a state-of-art problem, related to neural network training with using algorithms based on random search. One of the main problems while training using such algorithms is convergence problem. An original algorithm, based on modification of Boltzmann annealing scheme is proposed, for which theoretical convergence by probability to optimal solution from any initial is proved