[Clp] Performance optimization (Hauke Pribnow)

[Hauke Pribnow]

Tuomo and Gleb,

thanks for your suggestion. Since I'm completely new to this type of 
algorithms and operational research as a whole, I don't really 
understand your suggestion. Could you provide just a very small piece of 
example code that shows what you both mean?


By the way: I tested three free simplex solving libraries for 
performance in solving my pretty small problem. My test was set up to 10 
constant columns (where the associated variables had upper bounds) and 
one changing column (where the variable didn't have an upper bound).

I tested Clp (of course), Will Naylor's WNLIB (as suggested by William 
H. Patton) and LP_SOLVE.

My results are:
WNLIB took about 0.76310 seconds for 1000 changes in the last column and 
resolving the problem. Since WNLIB doesn't support upper bounds on 
variables natively, I had to introduce slack variables which even 
increased the size of the problem.

Clp took about 3.98016 seconds for the same test without the use slack 
variables of course. The same also goes for LP_SOLVE which took about 
3.14342 seconds.

Do you think that even if conditions like 'matrix is sparse', 'most 
columns are constant' or 'changes in the last column are pretty small 
most times' are met, Clp would outperform the other two libraries?

In case anyone is interested, I could clean up my code a bit and make it 
available for anyone who wants to take a look at it or also to do own tests.

Hauke