[R-pkgs] new version of minpack.lm
kate at few.vu.nl
Thu Mar 13 16:23:45 CET 2008
The package minpack.lm allows nonlinear regression problems to be
addressed with a modification of the Levenberg-Marquardt algorithm based
on the implementation of 'lmder' and 'lmdif' in MINPACK. Version 1.0-8 of
the package is now available on CRAN.
Changes in version 1.0-8 include:
o possibility to obtain standard error estimates on the parameters
via new methods for the generic functions 'summary' and 'vcov'
o possibility to extract other information via new methods for the
generic functions 'coef', 'deviance', 'df.residual', 'print',
o the argument 'control' of 'nls.lm' now defaults to
'nls.lm.control()'; 'nls.control.lm' allows a maximum number of
iterations to be specified; when the element 'nprint' of the
'control' argument of a call to 'nls.lm' is an integer greater
than 0, the residual sum of squares is now included in the
information printed every 'nprint' iterations
` o the list returned by 'nls.lm' includes elements 'niter' and
'deviance' that represent the number of iterations performed and
the residual sum of squares, respectively
side-note on Levenberg-Marquardt (LM) versus Gauss-Newton (GN):
There was some discussion
(http://finzi.psych.upenn.edu/R/Rhelp02a/archive/108758.html) on Rhelp
regarding whether one comes across real-world problems in which LM
performs better than GN. I have been seeing such problems recently in
some applications where GN as implemented in 'nls' reduces the step to a
very small value, resulting in little change in the residual sum of
squares from the starting values, whereas both NL2SOL applied via 'nls'
called with 'algorithm="port"' or LM as implemented in
'minpack.lm::nls.lm' significantly reduce the RSS. The implementation of
NL2SOL is slower by a significant factor on these problems as compared to
either the GN or LM implementations, making use of 'minpack.lm::nls.lm'
attractive. Note that these problems may be considered pathological;
there are issues with near collinearity of columns of the Jacobian and
with the assumption that the residuals are Gaussian.
More information about the R-packages