June 20, 2011

Success with building for windows!

June 19, 2011

Rude email from Ripley about not building on windows. Decided to
leverage the Dell netbook for this job.

June 18 2011, enroute from Zurich to Toronto

Lots of work on the demos, rgl, documents, and some code. Much tighter
now (docs consistent with code), better demos. The week in Uppsala was
devoted to the course and this package with perhaps 20-30 hours in
total devoted. The remaining issues concern building on windows... we
may be able to deal with this this week if all goes well, then two
articles, R journal and JSS and then the np function...

June 13 2011, Uppsala

So, the MESH parameters affect only continuous search since minimum
mesh for discrete is typically 1.

The parameters that affect discrete search are MAX_BB_EVAL and
INITIAL_MESH_SIZE.

June 11 2011, Uppsala

Looked deeper into which parameter value settings are driving
things. Noted that MAX_BB_EVAL was consistently being hit and search
terminating for the wage1 data set. Increased this to 10000 and it is
no longer binding... played with other parameter value settings and
modified INITIAL_MESH_SIZE but edited snomadr.Rd with some common
sense advice for setting values in crs

Also, MIN_POLL_SIZE is binding for mcycle (default is none?)

June 9 2011

En route to Uppsala and had a brain flash about using the lambda
approach and spitting a crs model with additional components in
crsiv... guess I might be able to use names of variables taken from
the data frames rather than x1, x2 etc.

June 7 2011

Merged snomadr (code for NOMAD is LGPL) into crs so now no
dependencies whatsoever (no gsl, no mgcv, no nomad) so it is
standalone. At this stage could upload to CRAN if I wished and sooner
rather than later I need to do so.

June 5 2011

Zhenghua has added a number of useful features.

One is odd - fit.z... I think include and lambda cover these. Will
meet tomorrow morning to discuss.

Also, need R_CheckUserInterrupt() to be called by snomadr as currently
cannot terminate...

Also, global scope variables need to be modified...

May 17, 2011, Toulouse

Zhenghua makes the big play. He has written snomadr, a standalone R
implmentation of NOMAD. Then he has written R functions that currently
smoke!

Dec 13 2010

Issue 1) I had been modifying endpoints by appending to knot
vector. This won't work. My previous `trimming' approach was in fact
better so perhaps I ought to revert to that.

Now, bs() in spline extends the endpoints (somehow) but throws a
warning, however, it does this without appending the knot vector. The
upshot is that the dimension is unchanged which is desirable - this is
all via the predict() function.

Issue 2) With quantile knots you can get repeated breakpoints when
data as point mass (rounded etc.). I had been reducing the number of
breakpoints dynamically when this happens, but this can throw things
off.

The bs() function handles this more gracefully.

I can envision hacks such as appending an epsilon to each identical
value and augmenting, but this is indeed a hack.

