Warm start best practice

[andrea-cassioli-maersk]

Hi, we are refactoring an old code based on clp that implements a column-generation algorithm.

The current implementation

• stores the master problem in a CoinModel object,
• accesses clp via OsiClpSolverInterface,
• stores the warm start information in a CoinWarmStart object
• uses the initialSolve the first time and then setWarmStart and resolve later on.

I am wondering if this is still the right way. I am looking at the ClpDualSimplex solver class and I am wondering whether under the hood it will performa warm start if I apply changes (basically adding columns) .

[jjhforrest]

Adding columns you would want ClpPrimalSimplex. Yes it will perform a warm start and the coding would be a bit simpler - but it would only be marginally faster. It may be worth looking at the example in Clp/examples - sprint.cpp. This is in many ways more complicated - but it may give you some ideas.

[andrea-cassioli-maersk]

Thanks @jjhforrest I will go for the ClpPrimalSimplex. For now simplicity is quite important over speed. The sprint.cpp example is a bit convoluted by I will look into it.

Btw, what is the meaning of the ifValuesPass parameter in ClpPrimalSimplex.primal()?