Comparison of three newton-like nonlinear least-squares methods for estimating parameters of ground-water flow models
Three methods of solving nonlinear least-squares problems were compared for robustness and efficiency using a series of hypothetical and field problems. A modified Gauss-Newton/full Newton hybrid method (MGN/FN) and an analogous method for which part of the Hessian matrix was replaced by a quasi-Newton approximation (MGN/QN) solved some of the problems with appreciably fewer iterations than required using only a modified Gauss-Newton (MGN) method. In these problems, model nonlinearity and a large variance for the observed data apparently caused MGN to converge more slowly than MGN/FN or MGN/QN after the sum of squared errors had almost stabilized. Other problems were solved as efficiently with MGN as with MGN/FN or MGN/QN. Because MGN/FN can require significantly more computer time per iteration and more computer storage for transient problems, it is less attractive for a general purpose algorithm than MGN/QN.
Citation Information
Publication Year | 1992 |
---|---|
Title | Comparison of three newton-like nonlinear least-squares methods for estimating parameters of ground-water flow models |
Authors | R.L. Cooley, M. C. Hill |
Publication Type | Conference Paper |
Publication Subtype | Conference Paper |
Index ID | 70016922 |
Record Source | USGS Publications Warehouse |