[R] ML optimization question--unidimensional unfolding scaling

[Peter Muhlberger]

Hi Spencer & Andy:  Thanks for your thoughtful input!  I did at one point
look at the optim() function & run debug on it (wasn't aware of
browser--that's helpful!).  My impression is that optim() simply calls a C
function that handles the maximization.  So if I want to break out of my
likelihood function to restart optim() w/ new values, it seems I'd have to
somehow communicate to C that it's time to stop.  May need to rewrite the C,
with which I'm not familiar--Java yes, so maybe when I have some real free
time....

Another possibility might be finding some jerry-rigged way to break out of
optim.  Maybe if I tell the likelihood function to freeze its returned value
at some point, optim will conclude it's done and stop.  Probably inefficient
& I will have the problem of telling when the break point ought to occur.
Just wish there were some programmatic way to say 'stop this and return
control to the higher-level calling function 'blah''.

A third possibility is one suggested by Spencer who seems to think it's ok
for the routine to pursue multiple branches w/o restarting, hence no restart
problem.  But w/ Newtonian-style convergence the latent scale values (which
are parameters to be estimated) have current positions & are supposed to
smoothly move toward lower likelihood values.  What will happen in branched
convergence, however, is that some of the latent values will prove to have
better values on the other side of a normal curve from their current
position.  My guess is that this will cause the likelihood function to make
a sudden, non-continuous jump not predictable by derivatives, which may mean
it can't converge properly.

Spencer's MDS alternative is intriguing & I'll need to think more about it.
Maybe I should also consider full-out Bayesian Monte Carlo methods (if I
have time), which would simultaneously explore the whole solution space.

Thanks,
Peter

On 11/2/05 9:01 PM, "Spencer Graves" <spencer.graves at pdf.com> wrote:

>  Have you looked at the code for "optim"?  If you execute "optim", it
> will list the code.  You can copy that into a script file and walk
> through it line by line to figure out what it does.  By doing this, you
> should be able to find a place in the iteration where you can test both
> branches of each bifurcation and pick one -- or keep a list of however
> many you want and follow them all more or less simultaneously, pruning
> the ones that seem too implausible.  Then you can alternate between a
> piece of the "optim" code, bifurcating and pruning, adjusting each and
> printing intermediate progress reports to help you understand what it's
> doing and how you might want to modify it.
> 
>  With a bit more effort, you can get the official source code with
> comments.  To do that, I think you go to "www.r-project.org" -> CRAN ->
> (select a local mirror) -> "Software:  R sources".  From there, just
> download "The latest release:  R-2.2.0.tar.gz".
> 
>  For more detailed help, I suggest you try to think of the simplest
> possible toy problem that still contains one of the issues you find most
> difficult.  Then send that to this list.  If readers can copy a few
> lines of R code from your email into R and try a couple of things in
> less than a minute, I think you might get more useful replies quicker.

On 11/3/05 8:08 AM, "Liaw, Andy" <andy_liaw at merck.com> wrote:

> Alternatively, just type debug(optim) before using it, then step through it
> by hitting enter repeatedly...
> 
> When you're done, do undebug(optim).

On 11/3/05 11:06 AM, "Liaw, Andy" <andy_liaw at merck.com> wrote:

> Essentially all that debug() does is like inserting browser() as the first
> line of the function being debug()ed.  You can type just about any command
> at the browser> prompt, e.g., for checking data, etc.  ?browser has list of
> special commands for the browser> prompt.
> 
> Andy