Branch-and-bound sub-MIPs pass near-zero (~1e-18) coefficients to Papilo

Aperta
#2,290 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
1-2 giorni
Idoneità per principianti
53/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva

Direzione di ricerca

Start with branch_and_bound_t::solve_submip and trace convert_lp_to_user_problem and third_party_presolve_t::apply_to_subproblem to see how coefficients reach Papilo. The issue points to cut row construction or sub-MIP conversion as possible places to drop tiny coefficients; compare against the stated tolerance context and verify that the pk1.mps sub-MIP no longer carries the near-zero values.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

bug mip

Branch-and-bound passes sub-MIPs to Papilo with coefficients around 1e-18.

Where

branch_and_bound_t::solve_submip builds the sub-MIP from worker->leaf_problem (structural rows, slacks and cuts) with convert_lp_to_user_problem, then calls third_party_presolve_t::apply_to_subproblem. Nothing drops tiny coefficients on the way.

Example

A sub-MIP of pk1.mps, captured at 3d71290b while running mip_solve.initial_solution_survives_papilo_crush. It has 54 rows; pk1 has 45, so the extra rows are most likely cuts. The last row, R53, has two coefficients on free columns:

  • C32: -1.734723475976807e-18 (exactly -2^-59)
  • C33: -1.894851796836205e-18

The other coefficients in that row are at least 0.015. A coefficient of exactly -2^-59 looks like leftover from floating-point cancellation rather than a real coefficient. Papilo's numerics.epsilon is 1e-9 and its presolve.minabscoeff is 1e-10.

Impact

Papilo's sparsify cancels C33 with a scale of -1.9e-18. That leaves every other coefficient in the row unchanged and triggers the Papilo assertion in scipopt/papilo#103, which #2267 patches. With the patch it no longer aborts, but the sub-MIP still carries values that are numerically zero.

Expected

Drop coefficients below a small tolerance when the cut rows are built, or when the sub-MIP is converted.

Lingua principale
Cuda
Stelle
1.1k
Fork
255
Merge medio
1g 22h
PR unite (30g)
102

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di NVIDIA/cuopt

Tutte le issue di NVIDIA/cuopt

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.