Nonsmooth, Nonconvex Optimization Using Functional Encoding and Component Transition Information
We have developed novel algorithms for optimizing nonsmooth, nonconvex functions in which the nonsmoothness is caused by nonsmooth operators presented in the analytical form of the objective. The algorithms are based on encoding the active branch of each nonsmooth operator such that the active smoot...
Gespeichert in:
Veröffentlicht in: | arXiv.org 2024-04 |
---|---|
1. Verfasser: | |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | We have developed novel algorithms for optimizing nonsmooth, nonconvex functions in which the nonsmoothness is caused by nonsmooth operators presented in the analytical form of the objective. The algorithms are based on encoding the active branch of each nonsmooth operator such that the active smooth component function and its code can be extracted at any given point, and the transition of the solution from one smooth piece to another can be detected via tracking the change of active branches of all the operators. This mechanism enables the possibility of collecting the information about the sequence of active component functions encountered in the previous iterations (i.e., the component transition information), and using it in the construction of a current local model or identification of a descent direction in a very economic and effective manner. Based on this novel idea, we have developed a trust-region method and a joint gradient descent method driven by the component information for optimizing the encodable piecewise-smooth, nonconvex functions. It has further been shown that the joint gradient descent method using a technique called proactive component function accessing can achieve a linear rate of convergence if a so-called multi-component Polyak-Lojasiewicz inequality and some other regularity conditions hold at a neighborhood of a local minimizer. |
---|---|
ISSN: | 2331-8422 |