gam allows the option to `polish' smoothing parameter
estimates by minimising the GCV/UBRE score calculated at convergence of the IRLS
algorithm, given a set of smoothing parameters, rather than using the `performance
iteration' method which estimates smoothing parameters within the IRLS loop. The
estimates are often slightly different, since the power iteration effectively
neglects the depedence of the iterative weights on the smoothing parameters.The `polishing' optimization is fairly crude and numerically costly.
nlm is used to minimize the scores with respect to the smoothin
parameters, and this routine is designed to be supplied to it as an argument.
This is basically a service routine for gam, and is not usually
called directly by users.
full.score(sp,G,family,control,gamma)GAMsetupgam.control"full.gam.object"
which is the full object returned by gam.fit.