Recursive RINS

[nguidotti]

This PR implements the recursive RINS heuristics similar to HiGHS. More specifically, it applies a dive-like procedure to create a RINS sub-MIP, presolved it with Papilo and then passes to the cut-and-branch procedure. It also launch a CPU FJ worker over the sub-MIP that runs in parallel to cut-and-branch.

Since the current implementation, does not have an unified problem representation, this PR also adds:

• A routine for converting a lp_problem_t (standard form) to user_problem_t (range form).
• Allow Papilo to be applied to an user_problem_t instead of optimization_problem_t.
• Adds an upper bound callback for retrieving the upper bound of the main solver (since it may change if another worker find a better incumbent).

Additionally, the number of worker per diving heuristic is no longer tied to the total number of workers, which allows the solver to use all diving heuristics in low thread counts (we rotate between the different heuristics for each diving worker launch).

Results

MIPLIB2017, 10min, GH200

================================================================================
 main-2026-07-03 (1) vs recursive-submip (2)
================================================================================

------------------------------------------------------------------------------------------------------------------------------
|                                        |       Run 1        |       Run 2        |     Abs. Diff.     |   Rel. Diff. (%)   |
------------------------------------------------------------------------------------------------------------------------------
| Imported                                                 240                  240                   +0                 --- |
| Feasible                                                 227                  228                   +1                 --- |
| Optimal                                                   86                   91                   +5                 --- |
| Solutions with <0.1% primal gap                          138                  151                  +13                 --- |
| Nodes explored (mean)                              1.326e+07            1.296e+07           -3.057e+05               -2.31 |
| Nodes explored (shifted geomean)                   1.262e+04            1.239e+04                 -230               -1.82 |
| Relative MIP gap (mean)                               0.2777               0.2726             -0.00512               -1.84 |
| Relative MIP gap (shifted geomean)                   0.09089              0.08901            -0.001883               -2.07 |
| Solve time (mean)                                      422.6                435.4               +12.79               +3.03 |
| Solve time (shifted geomean)                           196.9                226.6               +29.66               +15.1 |
| Primal gap (mean)                                      9.969                 9.29               -0.679               -6.81 |
| Primal gap (shifted geomean)                          0.4584               0.3491              -0.1093               -23.8 |
| Primal integral (mean)                                 25.02                22.97               -2.051                -8.2 |
| Primal integral (shifted geomean)                      2.849                2.421              -0.4275                 -15 |
------------------------------------------------------------------------------------------------------------------------------


----------------------------------------------------------------------
|             Name             |     status 1     |     status 2     |
----------------------------------------------------------------------
| neos-1171737                           feasible            optimal |
| radiationm40-10-02                     feasible            optimal |
| rail01                                  timeout           feasible |
| ran14x18-disj-8                        feasible            optimal |
| triptim1                               feasible            optimal |
| unitcal_7                              feasible            optimal |
----------------------------------------------------------------------

Checklist

• I am familiar with the Contributing Guidelines.
• Testing
  • New or existing tests cover these changes
  • Added tests
  • Created an issue to follow-up
  • NA
• Documentation
  • The documentation is up to date with these changes
  • Added new documentation
  • NA

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

🧹 Nitpick comments (1)

cpp/src/mip_heuristics/presolve/third_party_presolve.cpp (1)

1300-1307: 🚀 Performance & Scalability | 🔵 Trivial | ⚡ Quick win

get_coeff linear scan makes reduced-cost replay quadratic per row.

When the (row, col) key is absent from coeff_current (the common case), get_coeff scans the entire row in A. Callers iterate p over that same row and call get_coeff(row, A_indices[p]) (Lines 1354-1357 and 1441-1443), so each row costs O(row_nnz²). Since A_values[p] is already the coefficient for A_indices[p], only map-overridden entries need a lookup — pass the known value and consult the map only for overrides to keep the replay linear in nnz.

♻️ Illustrative change at the call sites

-      for (i_t p = A_offsets[i]; p < A_offsets[i + 1]; ++p) {
-        z[A_indices[p]] += y[i] * get_coeff(i, A_indices[p]);
-      }
+      for (i_t p = A_offsets[i]; p < A_offsets[i + 1]; ++p) {
+        const i_t col = A_indices[p];
+        auto it       = coeff_current.find(coeff_key(i, col));
+        const f_t a   = (it != coeff_current.end()) ? it->second : A_values[p];
+        z[col] += y[i] * a;
+      }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/src/mip_heuristics/presolve/third_party_presolve.cpp` around lines 1300 -
1307, The get_coeff helper currently does a linear scan of A for every miss in
coeff_current, which makes the reduced-cost replay quadratic when called inside
the row iteration loops. Update the logic around get_coeff and its call sites in
the replay path so the caller passes the already-known A_values[p] for
A_indices[p], and only consult coeff_current for overridden entries instead of
rescanning the row. Keep the fix localized to get_coeff and the loops that
invoke it so the replay stays linear in nnz.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@cpp/src/mip_heuristics/presolve/third_party_presolve.cpp`:
- Around line 1300-1307: The get_coeff helper currently does a linear scan of A
for every miss in coeff_current, which makes the reduced-cost replay quadratic
when called inside the row iteration loops. Update the logic around get_coeff
and its call sites in the replay path so the caller passes the already-known
A_values[p] for A_indices[p], and only consult coeff_current for overridden
entries instead of rescanning the row. Keep the fix localized to get_coeff and
the loops that invoke it so the replay stays linear in nnz.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: abae84ad-32f0-4d30-9d20-25a91ec0603b

📥 Commits

Reviewing files that changed from the base of the PR and between 869cc34 and 0a9f7ca.

📒 Files selected for processing (8)

• cpp/src/branch_and_bound/branch_and_bound.cpp
• cpp/src/branch_and_bound/constants.hpp
• cpp/src/branch_and_bound/pseudo_costs.hpp
• cpp/src/branch_and_bound/worker.hpp
• cpp/src/mip_heuristics/presolve/third_party_presolve.cpp
• cpp/tests/dual_simplex/unit_tests/solve.cpp
• cpp/tests/linear_programming/unit_tests/presolve_test.cu
• cpp/tests/mip/presolve_test.cu

🚧 Files skipped from review as they are similar to previous changes (7)

• cpp/tests/dual_simplex/unit_tests/solve.cpp
• cpp/src/branch_and_bound/constants.hpp
• cpp/tests/mip/presolve_test.cu
• cpp/src/branch_and_bound/pseudo_costs.hpp
• cpp/tests/linear_programming/unit_tests/presolve_test.cu
• cpp/src/branch_and_bound/worker.hpp
• cpp/src/branch_and_bound/branch_and_bound.cpp

[aliceb-nv]

Thanks a lot for the hard work and amazing results Nicolas! Approving with a few nits; let's also get Chris' eyes on the dual simplex/B&B parts

[aliceb-nv]

Let's check with Chris that these logging changes are agreed upon

[aliceb-nv]

Do we handle free rows properly in simplex?

[nguidotti]

It is a good question. Let me check, although @chris-maes or @rg20 may answer better than me.

[aliceb-nv]

Why do we pin on your repo? Is it because of a required upstream fix? If so, we probably should wait until it is merged before merging the PR instead of depending on a personal repo (or at least, track that with an issue)

[nguidotti]

Yeah, this is only temporary. I will wait for scipopt/papilo#89 to merged before merging this PR

[aliceb-nv]

The following was flagged by an analysis agent, do you believe it's legit?

At presolve.cpp:748-754, for a row with b = +inf and a standard [0, inf) slack:

• t1 = b − c·l_s = inf − 0 = inf
• t2 = b − c·u_s = inf − inf = NaN
• std::min(t1, t2) and std::max(t1, t2) both discard the NaN operand and return inf, so lo == hi == +inf
• the lo == hi branch fires → row recovered as sense 'E' with rhs = +inf — a vacuous constraint has become an unsatisfiable one.

The same collapse happens NaN-free if the row carries an artificial with bounds [0,0] (t1 = t2 = inf). Downstream, build_papilo_problem sets lhs = rhs = +inf with only the rhs infinity flag (third_party_presolve.cpp:304-315), so Papilo sees a^T x >= +inf and declares the sub-MIP INFEASIBLE. solve_submip records it via save_infeasible and moves on — every RINS/sub-MIP on such an instance silently reports infeasible, disabling the whole heuristic with no assert firing even in debug builds.

[nguidotti]

I do not think b can be inf here since $Ax = b$, which means that a row of the matrix has infinity coefficient or one entry in the solution vector x is infinite (this will be a weird problem hahaha). @chris-maes is this correct?

[coderabbitai]

Caution

Review failed

An error occurred during the review process. Please try again later.

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

This PR adds Sub-MIP/RINS branch-and-bound support, PaPILO-backed presolve crush/uncrush paths, an inverse simplex-to-user-problem conversion, simplex-iteration accounting updates, and related test and dataset changes.

Changes

Sub-MIP/RINS and Presolve Crush Support

Layer / File(s)	Summary
Settings, status, and worker contracts cpp/src/branch_and_bound/constants.hpp, cpp/src/branch_and_bound/deterministic_workers.hpp, cpp/src/branch_and_bound/worker.hpp, cpp/src/branch_and_bound/diving_heuristics.hpp, cpp/src/branch_and_bound/pseudo_costs.{hpp,cpp}, cpp/src/cuts/cuts.cpp, cpp/src/mip_heuristics/diversity/lns/rins.cu, cpp/src/mip_heuristics/diversity/recombiners/sub_mip.cuh, cpp/src/dual_simplex/simplex_solver_settings.hpp	Adds mip_status_t, SUBMIP, submip_settings_t, submip_stats_t, renames sub_mip to inside_submip and total_lp_iters to total_simplex_iters, adds warm-start pseudo-cost initialization, and updates diving-worker selection/state.
Presolver wrapper and inverse conversion cpp/src/dual_simplex/presolve.{cpp,hpp}, cpp/src/branch_and_bound/CMakeLists.txt, cpp/CMakeLists.txt, cpp/tests/dual_simplex/unit_tests/solve.cpp	Adds convert_simplex_problem to rebuild a range-form user_problem_t from simplex standard form, updates CMake wiring/formatting, and adds a round-trip conversion test.
PaPILO-backed crush/uncrush cpp/src/mip_heuristics/presolve/third_party_presolve.{cpp,hpp}, cpp/tests/mip/presolve_test.cu	Adds build_papilo_problem/build_user_problem overloads, a sub-MIP apply(...) presolve path, crush_primal_solution/crush_primal_dual_solution, and a presolve test asserting full reduction.
Branch-and-bound Sub-MIP/RINS pipeline cpp/src/branch_and_bound/branch_and_bound.{cpp,hpp}	Adds launch_submip_worker, solve_submip, rins, set_solution_from_submip, warm_start, apply_objective_step, external upper-bound callback support, and switches accounting to total_simplex_iters.
Diversity manager, early incumbents, and dataset updates cpp/src/mip_heuristics/diversity/diversity_manager.cu, datasets/mip/download_miplib_test_dataset.sh	Crushes user-given initial solutions via PaPILO, reorders initial-solution insertion, updates Sub-MIP flags, and extends the MIPLIB instance list.

Estimated code review effort: 5 (Critical) | ~120 minutes

Suggested reviewers: kaatish, hlinsen, Iroy30, KyleFromNVIDIA

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title matches the main change: recursive RINS heuristics and sub-MIP handling.
Description check	✅ Passed	The description accurately describes the PR’s RINS sub-MIP, Papilo presolve, upper-bound callback, and worker scheduling changes.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}