Repository navigation
Conversation
…t delta and actual cut separation
…nd the exact right hand side
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## latest #3473 +/- ##
==========================================
+ Coverage 75.85% 75.94% +0.09%
==========================================
Files 461 467 +6
Lines 115588 116368 +780
Branches 18059 18147 +88
==========================================
+ Hits 87677 88376 +699
- Misses 27671 27752 +81
Partials 240 240 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Opt-Mucca
left a comment
There was a problem hiding this comment.
I am a huge fan of these changes! The code is so much neater and my gut is telling that there's probably even a small performance improvement for cache efficiency. Great data structure additions.
I haven't approved yet because I'd like all the questions resolved first:
- Does the violation need to be measured in
HighsCDouble? I agree that we were occasionally getting cuts that disagreed with each other because of the different precision they were calculated in, but that's not really a correctness error? - Are we sure we want the "dummy cuts" to be calculated in
HighsCDoublewhen we're trying to findbestdelta? I assume this has a decent performance hit, and like the point above, it doesn't really fix any "true" correctness issues.
Thanks, @Opt-Mucca.
|
|
@fwesselm if upgrading to If it does: I always assumed that it was important to generate the dummy cuts using lower precision as well as the violation. As long as the final cut is computed using |
Description
@Opt-Mucca, these are the changes to simplify the code in
HighsCutGenerationandHighsTransformedLp. This has not been tested on a large-scale set yet:Row,TransformedVarandTransformedRowinHighsTransformedLp.h.flipComplementationused to computesolval = upper - solvalindouble. For integer variables with huge bounds, flipping back lost the solution value.HighsCutGeneration.cpp. Before, the search for the best scaling factor (delta) was computed indouble, while the cut was then built inHighsCDouble, and both version could disagree. Now, both use the same lambdas.Checklist
latestbranch