Hi Andrew, >> I am a bit more optimistic: Using the draft API, I saw some cases of >> ~20% improvement in pseudocost initialisation time. > > Did you use bfd_copy or the basis factorization was computed from > scratch for every re-optimization as currently implemented?
This was using bfd_copy in both cases, so the only change was using the internal API or not. Of course not every problem gained as much, but some (like triptim1 and app1-2 from miplib) did. The gain was calculated by counting the number of pseudocosts estimated at the root node, for the same fixed time interval in both cases. > Also please note that your changes might lead to some differences in the > pseudo-costs computed that, in turn, might affect the choice of > branching variables and thus change the overall solution time > signficantly. I tried to make the code produce the exact same results in every case, in order to keep the performance of the different approaches directly comparable, and in my tests this seems to be the case. The code would be a bit simpler if that was not the case but then the evaluation of the performance would be difficult, as you point out, and I'm not sure profiling would help there. Best Regards, Chris Matrakidis _______________________________________________ Help-glpk mailing list [email protected] https://lists.gnu.org/mailman/listinfo/help-glpk
