| gam.convergence {mgcv} | R Documentation |
When fitting GAMs there is a tradeoff between speed of
fitting and probability of fit convergence. The default fitting options,
specified by
gam arguments method and optimizer,
opt for certainty of convergence over speed of
fit. In the Generalized Additive Model case it means using
‘outer’ iteration in preference to ‘performance iteration’: see
gam.outer for details.
It is possible for the default ‘outer’ iteration to fail when finding intial
smoothing parameters using a few steps of performance iteration (if you get a
convergence failure message from magic when outer iterating, then this
is what has happened): lower outerPIsteps in gam.control
to fix this.
There are three things that you can try to speed up GAM fitting. (i) if you have large
numbers of smoothing parameters in the generalized case, then try the "bfgs" method
option in gam argument optimizer: this can be faster than the default. (ii) Change the
optimizer argument to gam so that ‘performance iteration’ is
used in place of the default outer iteration. Usually performance iteration
converges well and it can sometimes be quicker than the default outer iteration.
(iii) For large datasets it may be worth changing
the smoothing basis to use bs="cr" (see s for details)
for 1-d smooths, and to use te smooths in place of
s smooths for smooths of more than one variable. This is because
the default thin plate regression spline basis "tp" is costly to set up
for large datasets (much over 1000 data, say).
If the GAM estimation process fails to converge when using performance
iteration, then switch to outer iteration via the optimizer argument of
gam. If it still fails, try
increasing the number of IRLS iterations (see gam.control) or
perhaps experiment with the convergence tolerance.
If you still have problems, it's worth noting that a GAM is just a (penalized)
GLM and the IRLS scheme used to estimate GLMs is not guaranteed to
converge. Hence non convergence of a GAM may relate to a lack of stability in
the basic IRLS scheme. Therefore it is worth trying to establish whether the IRLS iterations
are capable of converging. To do this fit the problematic GAM with all smooth
terms specified with fx=TRUE so that the smoothing parameters are all
fixed at zero. If this ‘largest’ model can converge then, then the maintainer
would quite like to know about your problem! If it doesn't converge, then its
likely that your model is just too flexible for the IRLS process itself. Having tried
increasing maxit in gam.control, there are several other
possibilities for stabilizing the iteration. It is possible to try (i) setting lower bounds on the
smoothing parameters using the min.sp argument of gam: this may
or may not change the model being fitted; (ii)
reducing the flexibility of the model by reducing the basis dimensions
k in the specification of s and te model terms: this
obviously changes the model being fitted somewhat; (iii)
introduce a small regularization term into the fitting via the irls.reg
argument of gam.control: this option obviously changes the nature of
the fit somewhat, since parameter estimates are pulled towards zero by doing
this.
Usually, a major contributer to fitting difficulties is that the model is a very poor description of the data.
Simon N. Wood simon.wood@r-project.org