There have been several posts about optimizations where the parameters
for the objective function are bounds-constrained. Brian Ripley took my
1990 "Compact numerical methods for computers" codes and p2c'd
them to
give the CG and BFGS and (possibly, I should check!) the Nelder Mead
code. However, I have for use by myself and colleagues prepared variants
of the codes that allow bounds constraints and also what I call
"masks",
that is, fixed values of some parameters. The latter feature is often
useful when a parameter is usually fixed e.g., by legislation or
pegging, like an exchange rate, but may later be allowed to float. The
bounds are most helpful in forcing the user (us!) to think about scale.
I've also noticed they really do avoid some of the wild excursions
through parameter space and speed up convergence.
If anyone is interested in these possibilities, perhaps they could
contact me off-list and we can decide if there is sufficient willingness
to try to add the functionality in a graceful way. That is, we should
encourage but not require masks and bounds, and should make sure that
existing uses are not inconvenienced.
Cheers,
John Nash
