Software performance on nonlinear least-squares problems

Technical Report · Thu Sep 01 04:00:00 EDT 1988

OSTI ID:6716106

Fraley, C

This paper presents numerical results for a large and varied set of problems using software that is widely available and has undergone extensive testing. The algorithms implemented in this software include Newton-based linesearch and trust-region methods for unconstrained optimization, as well as Gauss-Newton, Levenberg-Marquardt, and special quasi-Newton methods for nonlinear least squares. Rather than give a critical assessment of the software itself, our original purpose was to use the best available software to compare the underlying algorithms, to identify classes of problems for each method on which the performance is either very good or very poor, and to provide benchmarks for future work in nonlinear least squares and unconstrained optimization. The variability in the results made it impossible to meet either of the first two goals; however, the results are significant as a step toward explaining why these aims are so difficult to achieve. 84 refs.

• Research Organization: Stanford Univ., CA (USA). Systems Optimization Lab.
• DOE Contract Number: FG03-87ER25030
• OSTI ID: 6716106
• Report Number(s): SOL-88-17; ON: DE89001867
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990210* -- Supercomputers-- (1987-1989)
990230 -- Mathematics & Mathematical Models-- (1987-1989)
ALGORITHMS
BENCHMARKS
COMPUTER ARCHITECTURE
ITERATIVE METHODS
LEAST SQUARE FIT
MATHEMATICAL LOGIC
MAXIMUM-LIKELIHOOD FIT
NEWTON METHOD
NONLINEAR PROBLEMS
NUMERICAL SOLUTION
OPTIMIZATION
PERFORMANCE
TABLES