[Scip] Is using a subset of SCIP (e.g. B&B) possible to extend Soplex capabilites for limited MIP functionality?

[Tobias Achterberg]

fergal mohan wrote:
> Hi guys, thanks for helping me tease out the details. I'm actually providing (Managed)
> C++ interface methods so that any .NET language can use the Solvers. I suspect that I
> can access the Soplex LP data (like dual variables, reduced costs or basis information)
> directly, even if I've used SCIP to setup the LP. In terms of the optimal design I'm
> inclined to start using SCIP for MIP and I'll see if I can make a conditional compile
> to build it eaither way (SCIP or Soplex only) for LP. I might be able to answer the
> overhead questions if all goes according to plan. Let me know if there is any more
> related wisdom on this, cheers, Fergal

This is great to hear that you are working on a .NET interface. Do you have plans to make
this public?

Regarding the LP solver interface, I don't think that mixing SCIP and Soplex stuff is a
good idea. They just have completely different interfaces and ways to store the problem
data. I would implement two separate wrappers, one for SCIP (with MIP capabilities) and
one for Soplex (with LP capabilities). Neither is a subset of the other. Soplex misses
integrality restrictions, SCIP is missing simplex information query methods like dual
solution values and basis information.

In this sense, SCIP+Soplex is really different from CPLEX. As you may know, in CPLEX you
can specify the problem with common interface methods, and it just depends on whether you
have a 'ctype' (integrality information) or not that decides whether the problem is an LP
or MIP. But similar to SCIP, if it is a MIP you cannot query duals or basis information
after having solved the problem.

As an example, consider a heuristic that solves a series of LPs and queries simplex
specific data. Eventually, it wants to declare some of the variables to be integer and
solve a MIP. This is very easy to do with CPLEX but hard to do with SCIP+Soplex. For
SCIP+Soplex, in order to solve a MIP you need to copy the whole problem from Soplex into SCIP.

Actually, such an algorithm in SCIP would better be implemented as a SCIP plugin, which
then uses SCIP's LP relaxation (which in the end will be a Soplex or any other LP solver
object) to fiddle around with LP stuff.


Tobias