A Preconditioned L-BFGS Algorithm with Application to Molecular Energy Minimization

Abstract

The limited-memory BFGS method has been widely used in large scale unconstrained optimization problems, such as the protein structure prediction problem. A major weakness of the L-BFGS method is that it may converge very slowly for ill-conditioned problems. We propose a preconditioned L-BFGS method, where we form the preconditioner from parts of the partially separable objective function. We report results of experiments in the context of the protein structure prediction problem for four different proteins, using a protein energy model as the objective function and multiple initial configurations for each protein. The results show speed-ups with factors between 3 and 10 in terms of function evaluations and with factors between 2 and 7 in terms of CPU time. The difference between CPU time and function evaluation speed-up is due to the extra overhead of calculating and applying the preconditioner. We also compare the performance of this method to the preconditioned truncated Newton method.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Nov 21, 2004
Accession Number
ADA444850

Entities

People

  • Elizabeth Eskow
  • Lianjun Jiang
  • Richard H. Byrd
  • Robert B. Schnabel

Organizations

  • University of Colorado Boulder

Tags

Communities of Interest

  • C4I

DTIC Thesaurus Topics

  • Abstracts
  • Algorithms
  • Availability
  • Classification
  • Colorado
  • Computer Science
  • Computers
  • Contracts
  • Heuristic Methods
  • Information Operations
  • Instructions
  • Mathematics
  • Monitoring
  • Optimization
  • Security
  • Test And Evaluation

Readers

  • Finite Element Method (FEM) for solving Partial Differential Equations (PDEs)
  • Operations Research