|
From: | Oscar Gustafsson |
Subject: | Re: [Help-glpk] Setting initial integer solution in GLPSOL |
Date: | Mon, 15 Dec 2008 11:58:35 +0100 (MET) |
On Mon, 15 Dec 2008, Andrew Makhorin wrote:
Another, related, feature would be the possibility to provide an initial bound from the command line. In many of the applications I work with (low-complexity FIR filter design) it is often pretty simple to find a sub-optimal solution, while solving for optimality is much harder. Providing the heuristic solution would of course be an alternative using Xyprons patch, but most of the time simply providing the objective function value would be enough and clearly easier.If you mean the incumbent objective value, it easy to specify it (at least in the mathprog model): just add a constraint like follows: s.t. foo: z <= bar; where z is the objective function (to be minimized), bar is its known upper bound.
Maybe I should clarify that I talk about (M)ILP problems, but that might have been clear from Xyprons mail.
I agree that this obviously is possible. However, at least for me it would be nice to have this command line option to keep the model "clean" (I do not always have the heuristic value, but when I do it would be nice to be able to simply utilize it). Also, wouldn't it be computationally more efficient to just initalize the B&B-solver cut-off rather than having an additional constraint to be processed?
I will run some examples to test it out more thoroughly as my initial test took longer to solve with the bound. However, as there is some "luck" involved in the B&B it is of course not possible to make any thorough claims yet.
Regards Oscar
[Prev in Thread] | Current Thread | [Next in Thread] |