Originally posted by: AnirudhSubramanyam

Hi,

I solve an MILP using cut and lazy callbacks. After finishing the solve, I change some of the separation routines that the callbacks use so that a different set of cuts/lazy constraints are now valid and applicable for the base lp object. Note that I do not change any coefficients, variables and/or constraints in the initial lp object. However, when I attempt to resolve, the second optimization spits out the exact same solution as the first one without even performing any kind of search. If I use the modified separation to begin with, then I get the solution that I expected (i.e., different from the solution corresponding to the original separation routines).

Is there a CPLEX function which allows me to "free" the tree information without having to call CPXXfreeprob() [as I would then have to re-create the "base MILP" model which I do not want to do, as model building consumes time] or CPXXcloneprob() [as I lose MIPstart information from the previous solves]?

Originally posted by: Laci Ladanyi

You are correct, cplex should reoptimize, but it does not :-(. I have registered a bug report. In the meantime a workaround is to use CPXXcloneprob(), but extract the mipstarts from the original MIP using CPXXgetnummipstarts() and CPXXgetmipstarts(), and then feed them to the cloned MIP using CPXXaddmipstarts().

Originally posted by: AnirudhSubramanyam

Thanks for the reply.

I tried implementing what you suggested in the following way. I have (original MILP, original separation); I call CPXXcloneprob() and I change my separation routine (via cbhandle information) without changing anything in the problem object and get (new MILP, new separation). I transfer my MIP starts from (original MILP, original separation) to (new MILP, new separation). I optimize (new MILP, new separation). I would now like to transfer all solution information from (new MILP, new separation) to (original MILP, original separation) and repeat the whole procedure in a loop. However, when I do this, the MIP starts from the original MILP are accepted in the new MILP without being filtered for feasibility through the new separation routine.

Is this also possibly a bug or is it intentional?

I finally implemented it using my own data structures for keeping track of new solutions found and transferring them to the cloned MILP. However, I have other applications where this is not always straightforward to do.