[R] Dealing with -Inf in a maximisation problem.

[Rolf Turner]

On 07/11/16 13:07, William Dunlap wrote:
> Have you tried reparameterizing, using logb (=log(b)) instead of b?

Uh, no.  I don't think that that makes any sense in my context.

The "b" values are probabilities and must satisfy a "sum-to-1" 
constraint.  To accommodate this constraint I re-parametrise via a 
"logistic" style parametrisation --- basically

    b_i = exp(z_i)/[sum_j exp(z_j)], j = 1, ... n

with the parameters that the optimiser works with being z_1, ..., 
z_{n-1} (and with z_n == 0 for identifiability).  The objective function 
is of the form sum_i(a_i * log(b_i)), so I transform back
from the z_i to the b_i in order calculate the value of the objective
function.  But when the z_i get moderately large-negative, the b_i 
become numerically 0 and then log(b_i) becomes -Inf.  And the optimiser 
falls over.

cheers,

Rolf

>
> Bill Dunlap
> TIBCO Software
> wdunlap tibco.com <http://tibco.com>
>
> On Sun, Nov 6, 2016 at 1:17 PM, Rolf Turner <r.turner at auckland.ac.nz
> <mailto:r.turner at auckland.ac.nz>> wrote:
>
>
>     I am trying to deal with a maximisation problem in which it is
>     possible for the objective function to (quite legitimately) return
>     the value -Inf, which causes the numerical optimisers that I have
>     tried to fall over.
>
>     The -Inf values arise from expressions of the form "a * log(b)",
>     with b = 0.  Under the *starting* values of the parameters, a must
>     equal equal 0 whenever b = 0, so we can legitimately say that a *
>     log(b) = 0 in these circumstances.  However as the maximisation
>     algorithm searches over parameters it is possible for b to take the
>     value 0 for values of
>     a that are strictly positive.  (The values of "a" do not change during
>     this search, although they *do* change between "successive searches".)
>
>     Clearly if one is *maximising* the objective then -Inf is not a value of
>     particular interest, and we should be able to "move away".  But the
>     optimising function just stops.
>
>     It is also clear that "moving away" is not a simple task; you can't
>     estimate a gradient or Hessian at a point where the function value
>     is -Inf.
>
>     Can anyone suggest a way out of this dilemma, perhaps an optimiser
>     that is equipped to cope with -Inf values in some sneaky way?
>
>     Various ad hoc kludges spring to mind, but they all seem to be
>     fraught with peril.
>
>     I have tried changing the value returned by the objective function from
>     "v" to exp(v) --- which maps -Inf to 0, which is nice and finite.
>     However this seemed to flatten out the objective surface too much,
>     and the search stalled at the 0 value, which is the antithesis of
>     optimal.
>
>     The problem arises in a context of applying the EM algorithm where
>     the M-step cannot be carried out explicitly, whence numerical
>     optimisation.
>     I can give more detail if anyone thinks that it could be relevant.
>
>     I would appreciate advice from younger and wiser heads! :-)
>
>     cheers,
>
>     Rolf Turner
>
>     --
>     Technical Editor ANZJS
>     Department of Statistics
>     University of Auckland
>     Phone: +64-9-373-7599 ext. 88276 <tel:%2B64-9-373-7599%20ext.%2088276>
>
>     ______________________________________________
>     R-help at r-project.org <mailto:R-help at r-project.org> mailing list --
>     To UNSUBSCRIBE and more, see
>     https://stat.ethz.ch/mailman/listinfo/r-help
>     <https://stat.ethz.ch/mailman/listinfo/r-help>
>     PLEASE do read the posting guide
>     http://www.R-project.org/posting-guide.html
>     <http://www.R-project.org/posting-guide.html>
>     and provide commented, minimal, self-contained, reproducible code.
>
>


-- 
Technical Editor ANZJS
Department of Statistics
University of Auckland
Phone: +64-9-373-7599 ext. 88276