Extending Accelerated Optimization into the PDE Framework
Abstract
The primary goal is to extend momentum based methods (Heavy Ball, and Nesterov) from the finite dimensional framework into the infinite dimensional framework of gradient based optimization over the manifolds of functions, curves, surfaces, diffeomorphisms, and other distributed unknowns. This includes developing 1) a theory of PDE methods for accelerated gradient optimization 2) formulation of the specific PDE equations for the application of this theory in several optimization contexts 3) numerical methods for solving accelerated PDE systems 4) theoretical analysis of convergence rates and performance gains 5) development of efficient implementations amenable to parallel computing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 23, 2021
- Accession Number
- AD1197255
Entities
People
- Anthony Yezzi
Organizations
- Georgia Tech Research Corporation