Subscribe to Dr. Granville's Weekly Digest

Hi,

My company has asked me to pick a optimization solver for my project which involves solving a large MIP problem.  I have narrowed the search down to two vendors: IBM ILOG CPLEX vs FICO XPRESS. 

 

CPLEX has been the industry's gold standard for years but XPRESS put up a good argument that some of the big companies have abandoned CPLEX for XPRESS for a better solver performance and more flexible modeling environment.  Could someone share with me their experience on this subject?  Thanks.

Views: 8634

Replies to This Discussion

Hi, Michael:
Thanks. I have also talked to Gurobi team and as far as I know, their founding memebers came from ILOG CPLEX before IBM acquired it. I have no doubt about Gurobi but the problem for me is that now I have three companies each of which is telling me they are the best. I need experts, like yourself, to tell me pros and cons of these three solvers. Thanks.
I've used both to solve large problems, and between the two, i would get price estimates for how you would need to configure your licenses (i know Xpress starts cheap but as you add features it get pretty comparable to CPLEX last i checked--several years ago) and go with the cheaper option. Both are very fast, stable, and have plenty of really nifty modeling language feature that make implementing performance enhancing algorithms easy.

But i will also suggest you investigate the solvers and modeling languages maintained here:
http://www.coin-or.org/

None of these solvers have the spiffy interfaces, and they will require much heavier programming and intimate knowledge of how to implement advanced MIP algorithms, but they are open source and free. the only draw back is that you don't have the might of a corporate user support engine.

Best,
Jonathan
Hi, Jonathan:
Thanks. I got price estimates from both companies and each one is willing to match the other's quote. How do you feel about their technical support and consulting services?

I am the only one in my company with optimization background and hence if I have any modeling issues, I have no one to turn to but the company itself.
I would definitely look at COIN-OR as a possibility. Also look at GLPK. I've used GLPK but not on very large datasets. Although it is rather configurable (if you are knowledgable in C).

As to the question of user support. Although there is no corporate support there is the community support. I would argue that community support is much better than corporate support.

Pros open source support
1) Community support is not biased to software, methods, or technical requirements
2) Much larger support base
3) Much larger breadth of user experience
4) Many different contact channels (web forums, IRC chat, mailing lists, email)

Cons open source support
1) Not immediate but pretty darn close
2) Not guaranteed so mileage may vary
Hi Yi-Chun!

The best option is to look at the alternatives yourself. Figure out which option gives you the right combination of speed, features, price and service. The vendors offer various evaluation copies and benchmarking services that can help you decide.

Disclaimer: I currently work for Gurobi and formerly worked for ILOG-CPLEX.
Hi, Greg:
I saw the benchmark report and since you have also worked at ILOG, can you tell me why Gurobi is outperforming CPLEX?
We're pleased with the public benchmarks, but they should be "taken with a grain of salt". The benchmark that matters most is how well does a solver work on the models that you want to solve, as opposed to a standard academic test set. That's why you should test the different solvers and see which one best meets your needs.

As for why Gurobi did well in these public benchmarks, the best thing I can say is that we're learning from our experience with ILOG-CPLEX.
Hi, Greg:
Can you be more specific about what you learned with cplex? Is there any technical advantage that Gurobi has over CPLEX?
I can't get into details, but I can say that Gurobi was designed to be streamlined and efficient. Both the simplex and the MIP solvers are designed to be as fast as possible.
Thanks Greg. Can you talk a little about XPRESS? How is their solving compared to CPLEX and/or Gurobi? Thanks.
I have no first-hand experience with Xpress, so I'm not qualified to comment on it. However, it's interesting to note that they no longer participate in the public benchmarks by Prof. Hans Mittelmann.
One more thought on benchmarks: pay attention to the parameter settings. The major solvers let you set parameters to improve the performance; a few changes can make a nice improvement in your solve times. To get a fair comparison, don't compare the default settings from one solver with the recommended settings you get from one vendor. (On the other hand, be wary if anyone tells you to change a long list of parameters - you don't want to over-tune the results to a small set of test models).

RSS

© 2014   AnalyticBridge.com is a subsidiary and dedicated channel of Data Science Central LLC

Badges  |  Report an Issue  |  Terms of Service