Nadia Udler
University of Connecticut (Stamford)

Abstract

Machine learning models are often represented by functions given by computer programs. Optimization of such functions is a challenging task because traditional derivative based optimization methods with guaranteed convergence properties cannot be used.. This software allows to create new optimization methods with desired properties, based on basic modules. These basic modules are designed in accordance with approach for constructing global optimization methods based on potential theory KAP. These methods do not use derivatives of objective function and as a result work with nondifferentiable functions (or functions given by computer programs, or black box functions), but have guaranteed convergence. The software helps to understand principles of learning algorithms. This software may be used by researchers to design their own variations or hybrids of known heuristic optimization methods. It may be used by students to understand how known heuristic optimization methods work and how certain parameters affect the behavior of the method.

Keywords

global optimization, black-box functions, algorithmically defined functions, potential functions

DOI

10.25080/majora-212e5952-019

Bibtex entry

Full text PDF

Copyright © 2022 Nadia Udler. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.