Keeping a horse in the race, 2014 edition
This year’s conference is a huge success. More than 5000 participants, so many events, and so many talks! A huge proportion of those talks propose new or improved solution methods for mixed-integer programs. And the most popular way to demonstrate the interest of these methods is to benchmark against a generic solver. Last year, I heard John Siirola called this keeping a horse in the race. I very much liked the expression.
Over the last year, MIP solvers have improved. Often to the point that many researchers are close to being beaten by the horse they put in the race. For example, I attended some talks in the stochastic programming track. The authors have implemented a state-of-the-art decomposition algorithm (including acceleration techniques) for a multistage stochastic programming problem. Yet, the MIP solver they compare against still beats them for about 20-30% of instances. Solvers are not only catching up with applications, in some fields they are pretty much catching up with tailored algorithms from academia.
There are many possible reasons for this. First is that many ideas that get into solvers these days do not come from research papers anymore. Availability of multi-core processors is also a factor, and I assume the typical code from academia doesn’t match the level of sophistication that is found in code-optimized commercial solvers. They also have access to a huge model base to test their new ideas which is considerably richer than the typical 5 to 30 instances found in a research paper. With all the new features presented at this Conference, I imagine next year’s horses will be even more challenging to beat!