Page 1 of 1

CPLEX LP Warmstart

Posted: Tue Feb 14, 2017 9:59 pm
by lego
Hi all,

I'm solving a huge LP using CPLEX. I know that the final solution will be quiet similar to a solution, that I can obtain very easily, if I ignore a couple of equations. So I thought I could speed up things, if I solve the problem without the named equations and then pass the solution as a starting point to the actual problem.

To test working with an initial solution I solved the whole LP on a small scale, saved the solution and set all variables to the optimal solution before starting it again, but the number of necessary iterations won't decrease. I tried different algorithms for solving without any change. From my understanding at least the Simplex algorithm should benefit from an optimal initial solution. I've looked closely at the CPLEX options the GAMS documentation provides without any success.

I'm thankful for every support you can provide.

Regards,

Leo

Re: CPLEX LP Warmstart

Posted: Wed Feb 15, 2017 8:47 am
by bussieck
Leo,

The key GAMS/Cplex option for LP restarts is 'advind' (http://www.gams.com/latest/docs/solvers ... PLEXadvind). Perhaps you can post some logs or even share your model and starting point, so we have more to go on.

-Michael

Re: CPLEX LP Warmstart

Posted: Wed Feb 15, 2017 9:52 pm
by lego
Hey Michael,

first of all thanks for your answer. I managed to embed 'advind' successful, which is a huge improvement. However can it only be used, if there is a preexisting basis from a previous solve statement in the same code? I would like to generate the initial solution in a another GAMS file or even outside of GAMS and then import it via GDX.

I thought about a dummy problem without a real object function in order to create a Basis from a initial solution, but didn't succeed. I added both the model and the logfile.

Re: CPLEX LP Warmstart

Posted: Thu Feb 16, 2017 3:42 pm
by bussieck
It does not matter where the basis (complete or incomplete) comes from. You just need to set (.l and) .m of your variables and equations. I did not get the experiment you tried to do in your attached GAMS model. To me it looks like that the second solve successfully restarted from a basis.

-Michael