[Clp] What contributions would improve MIP performance?

[Bannister, Hugh]

Hi Keith
 
Right now get into trying to improve the performance of any LP or MlP solvers (other than by tuning the solver parameters against our problem).  That's a big task with limited chance of success and we don't have enough in the bank for that sport of speculation.  The sensible course is to test and improve our model formulation and solve strategies.  By all means test that against different optimisers.  Please focus on that for now and we can debate other approaches when I get back.
 
Please copy Ken into this type of correspondence.  Piecewise linear is not sufficient for us; we are dealing with non-convexity.  Ken knows all about this.  I worry we are tending to run off tackling things that are not likely to go anywhere and are not focussing enough on things that will count.
 
Hugh

________________________________

rom: Bannister, Keith
Sent: Thu 13/03/2008 4:15 PM
To: clp at list.coin-or.org
Cc: de Jong, Ron; O'Brien, Ricky; Bannister, Hugh
Subject: What contributions would improve MIP performance?



Hi There!

 

We're trialing CBC against a few other solvers against a wide variety of problems, the largest of which is 350k elements, 65k constraints, 500 binary variables. Just the relaxed problem takes CLC 32s to solve, but XPRESS only 9s. Solving the MIP problems often takes hours but I haven't got any side-by-side comparison data yet.

 

So, we're trying to improve performance - I have one offer, and one question.

 

Offer:

If I had 2 weeks to spend - I'm not a gun programmer but I can have a go at some ideas - which areas would you suggest I can help improve performance? 

 

I was thinking of linking in the Intel Math Kernel Library http://www.intel.com/cd/software/products/asmo-na/eng/307757.htm into CLC but I'm not sure whether CLC really spends that much time doing cholesky, matrix mul & other numerical things, does it? Is calling BLAS/LAPACK style functions a bit taboo these days? 

 

Can anyone suggest any ideas to improve performance in CLC/CBC that you would implement if only they had the time?

 

Question:

The piecewise linear feature would suit a lot of aspects of our model for which we would otherwise use integer or SOS1. Before we re-formulate everything, should piecewise linear on say 7 pieces generally be more efficient than integer or SOS1?

 

All the best and good work.

 

Keith Bannister

Intelligent Energy Systems

 


*******************************************************************************

This e-mail (which includes any attachments) is intended only for the use of the individual or entity named above and may contain information that is confidential and privileged and the copyright of Intelligent Energy Systems Pty Ltd or a third party. Any views expressed in this e-mail are those of the individual sender, except where the sender specifically states them to be the views of Intelligent Energy Systems Pty Ltd. There is no warranty that this e-mail is error or virus free. If you are not the intended recipient, you must not disclose or use the information contained in it. If you have received this e-mail in error, please notify the sender immediately by return e-mail and delete the original message.



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/clp/attachments/20080313/5bc7e56c/attachment.html>