Heuristic Improvements with solution hash, MAB and simplex root solution #216

akifcorduk · 2025-07-17T11:38:09Z

Description

This PR adds improvements to heuristics framework with the following:

Adds solution hash to check if an integer assignment has been locally searched before. If so, we skip that after few trials.
Adds dual simplex root relaxation solution to heuristics and halts(or prevents) PDLP root run.
Implements the multi armed bandit algorithm for the LS, currently disabled because no clear benefit was observed. The experiments will be continued here.
Generalized the MAB algorithm so that it could be used for any purpose which has discrete options to choose from.

…stic_improvements

copy-pr-bot · 2025-07-17T11:38:13Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

akifcorduk · 2025-07-17T11:39:15Z

/ok to test 03ade15

aliceb-nv

Looking good :) Just a few questions regarding the hashing
Thank you Akif!

cpp/src/mip/diversity/assignment_hash_map.cu

aliceb-nv · 2025-07-17T12:46:25Z

cpp/src/mip/diversity/assignment_hash_map.cu

+#if !MIP_INSTANTIATE_FLOAT && !MIP_INSTANTIATE_DOUBLE
+static_assert(false, "MIP_INSTANTIATE_FLOAT or MIP_INSTANTIATE_DOUBLE must be defined");
+#endif


Those atrocious linking errors :) I wish there was a way to throw an error at preprocessor/link time without having to add explicit checks

Yeah, this is the second time i spend significant amount of time to these issues.

Should we just remove FLOAT? do we anticipate using it for MIP?

aliceb-nv · 2025-07-17T13:02:14Z

cpp/src/mip/diversity/assignment_hash_map.cu

+  thrust::gather(solution.handle_ptr->get_thrust_policy(),
+                 solution.problem_ptr->integer_indices.begin(),
+                 solution.problem_ptr->integer_indices.end(),
+                 solution.assignment.begin(),
+                 integer_assignment.begin());


I think we ought to be careful with the implicit cast to size_t performed here. The C++ standard states that any value that cannot be represented in the target cast type yields undefined behavior:

Real floating-integer conversions
A finite value of any real floating type can be implicitly converted to any integer type. Except where covered by boolean conversion above, the rules are:

The fractional part is discarded (truncated towards zero).
If the resulting value can be represented by the target type, that value is used
otherwise, the behavior is undefined.
int n = 3.14; // n == 3
int x = 1e10; // undefined behavior for 32-bit int

I think this mean there might be issues in the case of negative and large (>1e20) assignments. It's probably better to compute the hash in a bitwise manner

Good catch, fixed!

aliceb-nv · 2025-07-17T13:54:57Z

cpp/src/mip/diversity/assignment_hash_map.cu

+    hash_1 ^= hash_2 + 0x9e3779b9 + (hash_1 << 6) + (hash_1 >> 2);
+    return hash_1;


I think this is a 32-bit hash, with 64-bit values this might not work properly

Thanks fixed!

cpp/src/mip/diversity/diversity_manager.cu

akifcorduk · 2025-07-18T09:27:00Z

/ok to test 6bcd726

akifcorduk · 2025-07-18T10:57:25Z

/ok to test 08d8372

aliceb-nv

Awesome work Akif!

aliceb-nv · 2025-07-18T13:22:23Z

cpp/src/mip/diversity/assignment_hash_map.cu

                 solution.problem_ptr->integer_indices.end(),
                 solution.assignment.begin(),
-                 integer_assignment.begin());
+                 reinterpret_cast<double*>(integer_assignment.data()));


I seriously doubt we will ever use anything but FP64, but in the rare case we instantiate the solver with f_t=FP32, this is going to break/behave badly. We might want to add a static_assert to make sure sizeof(f_t) == sizeof(size_t) in order not to cause issues if users decide to try out FP32

okay let me add the static assert here.

akifcorduk · 2025-07-18T13:45:55Z

/ok to test a92b9d4

akifcorduk · 2025-07-22T07:51:22Z

/merge

akifcorduk added 30 commits July 1, 2025 04:49

first merge: merging only the mip directory

1bf2309

add more changes from include directory

d49d073

fix merge conflicts

542e4b4

fix merge conflicts

6af2454

fix more compile errors

24cbd7c

remove patch files

18bfa3c

remove patch files

804d3d5

remove more unnecessary code

4040daf

log to console by default

345cc5e

add default lp state for fixed and normal problem

ff42e2c

add var_t header to host_helpers

f77e436

Merge branch 'branch-25.08' of github.com:NVIDIA/cuopt into new_heuri…

3200fb8

…stic_improvements

implement MAB for LS and fix a bug

d485ca8

fix few bugs and add checks

404763c

add missing header

d2a704e

fix bugs in MAB and improve the normalization logic

8207b2d

higher recency

4559128

revert majority of the settings without the LS MAB

bed3ea2

with LS MAB

957eccd

fix logging with LS MAB

ed5e347

without LS MAB

be7bfb9

revert more mab changes

53a802d

revert logging changes

0a84c94

by default check for both iter and time limit

f3457a1

enable mab for LS

e1ab729

disable mab ls

a8a2fef

implement solution hash check in local search

00a51c8

lower tolerance

767828d

add dual simplex root relaxation to heuristics

9d3c128

implement stopping of root pdlp if simplex is found

55bc1e2

akifcorduk added 2 commits July 17, 2025 04:14

Merge branch 'branch-25.08' of github.com:NVIDIA/cuopt into mab_for_ls

08aab8e

fix few merge conflicts

03ade15

akifcorduk added this to the 25.08 milestone Jul 17, 2025

akifcorduk requested review from a team as code owners July 17, 2025 11:38

akifcorduk requested a review from tmckayus July 17, 2025 11:38

akifcorduk added non-breaking Introduces a non-breaking change improvement Improves an existing functionality labels Jul 17, 2025

akifcorduk requested review from aliceb-nv and rg20 July 17, 2025 11:38

akifcorduk added the mip label Jul 17, 2025

akifcorduk added 2 commits July 17, 2025 05:45

use set device in case there are other gpus in the system

2adeb6b

fix user bound issues

aceb680

aliceb-nv reviewed Jul 17, 2025

View reviewed changes

rgsl888prabhu approved these changes Jul 17, 2025

View reviewed changes

rgsl888prabhu and others added 3 commits July 17, 2025 11:43

Merge branch 'branch-25.08' into mab_for_ls

d33c4d8

handle review comments

9074349

Merge branch 'mab_for_ls' of github.com:akifcorduk/cuopt into mab_for_ls

6bcd726

add objective check

08d8372

aliceb-nv approved these changes Jul 18, 2025

View reviewed changes

akifcorduk added 2 commits July 18, 2025 06:30

add static assert for double precision

22ac5a5

disable the batch python tests

a92b9d4

rapids-bot bot merged commit 47fb158 into NVIDIA:branch-25.08 Jul 22, 2025
344 of 352 checks passed

		hash_1 ^= hash_2 + 0x9e3779b9 + (hash_1 << 6) + (hash_1 >> 2);
		return hash_1;

Heuristic Improvements with solution hash, MAB and simplex root solution #216

Heuristic Improvements with solution hash, MAB and simplex root solution #216

Uh oh!

Conversation

akifcorduk commented Jul 17, 2025

Description

Uh oh!

copy-pr-bot bot commented Jul 17, 2025

Uh oh!

akifcorduk commented Jul 17, 2025

Uh oh!

aliceb-nv left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

akifcorduk commented Jul 18, 2025

Uh oh!

akifcorduk commented Jul 18, 2025

Uh oh!

aliceb-nv left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

akifcorduk commented Jul 18, 2025

Uh oh!

akifcorduk commented Jul 22, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants