[SCIP] scip + clp ?

[Marc Pfetsch]

On 30/11/17 01:25, Stuart Rogers wrote:
> This documentation (http://scip.zib.de/doc/html/LPI.php
> <http://scip.zib.de/doc/html/LPI.php>) states that the SCIP interface to
> the LP solver CLP sometimes produces wrong results. How serious is this
> issue and why hasn't it been addressed?

I you run the small testset that comes along with SCIP you will
recognize that there are instances for which SCIP with Clp computes
wrong "optimal" solutions (i.e., they are suboptimal). I would say that
this is serious, but you have to judge for yourself.

Why hasn't this been addressed? Well, I tried to debug this, but it is
very hard to track these problems down. On the mentioned instances, they
do not appear for a single LP, but only if the LPs are solved in
sequence. This is where I stopped. However, you would be very welcome if
you help to find the problem.

> CLP is free, yet performs just
> as well as the expensive LP solver CPLEX and better than the LP solver
> SoPlex according to these benchmarks
> (http://plato.asu.edu/ftp/lpsimp.html
> <http://plato.asu.edu/ftp/lpsimp.html>). In all my tests so far on my
> particular model, setting the LP solver to CLP makes SCIP terminate very
> rapidly to a very non-optimal solution.

Yes, Clp is a fast and free solver, especially on the mentioned
benchmarks. Note, however, that it might not be faster if you apply it
within branch-and-bound. In the meantime, I would expect SoPlex to be
faster that Clp on average, but I have not tested this for a long time
because of the above mentioned problems.

Best

Marc