Repository navigation
[FEA] Determinism for MILP heuristics #144
Description
Scope for 26.02:
- Improve work limits in Dual Simplex
- First working version of phase2 in dual simplex
- Deterministic Parallel Branch and Bound without heuristics
- Deterministic exchanges/sync between Heuristics + Branch and Bound
- Minimal working version
- Tuned for performance
Activity
anandhkb commentedon Aug 11, 2025
@aliceb-nv
Please call out your acceptance criteria for 25.10. Merci!
@afender for viz
aliceb-nv commentedon Aug 11, 2025
Acceptance criteria:
In 25.10, consecutive solver runs in MIP heuristics mode over the MIPLIB2017 benchmark set performed with a given work limit MUST see their objective value per instance remain within a certain epsilon of the corresponding objective value of previous runs. This epsilon remains to determine (<0.01% rel err?).
How to test: perform N (to determine? 5?) runs back-to-back over MIPLIB2017 and heuristics-only=t, with work_limit=W and seed=SEED. The Linf distance between the objective relative error vectors of each pair of runs must remain < rel_err.
anandhkb commentedon Feb 10, 2026
@aliceb-nv Your Determinism 26.02 PRs have been merged. What do you advise?
aliceb-nv commentedon Feb 10, 2026
We still need to combine the deterministic GPU heuristics with the deterministic parallel B&B. This will be the goal of this release (26.04)
anandhkb commentedon Feb 10, 2026
TBD for 26.04
Combine the deterministic GPU heuristics with the deterministic parallel B&B