Originally posted by: excalibur1491

Hi,

I am solving a pretty huge problem using MIP, and since it might be hard to get optimal solutions, we are thinking of taking a Local Search or Large Neighbourhood Search approach. 

The difficulty of the problem comes from having a bunch of Binaries, everything else are continuous variables.

So I have the following two questions:

-To do an LNS, I first assign all the binaries and find a solution. I then undo some of the decisions (using CPXchgbds to reset the binaries to 0 <= b <= 1) and resolve. I will be doing this over and over.  Is there a specifically good way of doing this in CPLEX? The way i am thinking of doing it is by using CPXchgbds, but is there a better way that allows CPLEX to use things it already did in previous iterations?

-To do a Local Search, I would assign all the binaries, solve the remaining LP (once the binaries are assigned, the problem is just an LP). Then switch a couple of binaries (say I set one that was True to False and one that was False to True). Then I get another LP that I would like to solve. Knowing that what CPLEX will be solving are LPs (and not MIPs as above), is there a specifically good way of doing this so CPLEX can re-use previous work?

Thanks!

Originally posted by: BoJensen

Using using CPXchgbds () is perfectly fine, CPLEX will reuse the stored solution automatically.

When repotimizing your problem should be considered a continuous problem, there's a convenient way of doing this :

http://www.ibm.com/support/knowledgecenter/SSSA5P_12.6.3/ilog.odms.cplex.help/CPLEX/UsrMan/topics/discr_optim/mip/para/51_soln_fixed.html

Fixing all the integer/binary variables sounds more like a rounding heuristic than a LNS to me.
 

Originally posted by: excalibur1491

Excellent, thanks for your reply.

Just one thing: you say fixing all the integer variables is not an LNS. But the LNS I had in mind was to fix them all in order to get a trivial initial solution, then unfix a few, reoptimize that few, unfix another few, reoptimize that few etc.
 Isn't that what LNS is about?

Originally posted by: BoJensen

Yes that would be a LNS approach.