[Ipopt] Solving LPs

[David Veerasingam]

> The easiest method to find an efficient solver for your problem is to
> produce your problem in a standard format such as MPS and test it on
> evaluation versions of Mosek and Gurobi.  Mosek in particular might be
> a good candidate to test.

I just wanted to echo the opinion that it is probably best to use a  
strictly LP code for solving LPs/network problems. This is especially  
important if one wants to compare the results with other methods. In  
opinion, using a piece of NLP code (even a superb work of art like  
IPOPT, of which I'm a strong advocate) to solve LPs may not give fair  
and valid comparisons to other solution methods.

Also, I want to point out that if you're in an academic institution,  
Gurobi and CPLEX (full versions) are free, as in gratis. They both  
come with binaries that will work in AMPL. You can get CPLEX 12.2  
under the IBM academic initiative, and Gurobi can be obtained by  
signing up on their website. Both these solvers have special  
heuristics for LPs, and can take advantage of multiple cores on your  
computer. They also both implement simplex and barrier (interior- 
point) methods. CPLEX in particular has a automatic tuning mode, in  
which it analyzes your problem (heuristically) and spits out runtime  
options that could potentially lead to faster solutions. I've tried  
this mode and in some cases, the right CPLEX options have improved my  
solve times by more than 20%.

David