Skip to content

Rank solutions with Space::compare in best-solution search - #242

Draft
zayenz wants to merge 21 commits into
mainfrom
feature/comparison
Draft

zayenz wants to merge 21 commits into
mainfrom
feature/comparison

Conversation

@zayenz

@zayenz zayenz commented Sep 10, 2026 •

Copy link
Copy Markdown
Member

Adds a read-only Space::compare() that ranks two solutions as better, equivalent, worse, or incomparable. Every space used for optimization must now implement both constrain() and compare(). constrain() still posts the improvement restriction; compare() decides which of two solutions to keep. The default compare() throws SpaceNoComparison.

Engines:

  • Sequential BAB and RBS: when an external solution arrives through constrain, compare() decides whether it replaces the incumbent.
  • Parallel BAB and PBS portfolios: every solution is compared with the retained incumbent. Only SpaceComparison::Better replaces it, so equivalent and worse solutions are dropped. If compare() throws, the exception reaches the caller, and the engine has to be deleted. A stored worker failure is reported before any further comparison, so the original exception is never replaced by a later one.
  • A* and CABS (experimental): compare() decides which solutions become the incumbent. The guidance bound still orders and prunes open nodes.
  • All engines reject incomparable solutions with Search::Incomparable.

Models and helpers:

  • SpaceComparison is an enum class (Better, Equivalent, Worse, Incomparable). Gecode::Sense (Minimize, Maximize) is a shared optimization direction.
  • compare_int, compare_float and compare_lex compare integer, float and lexicographic objective values for a Sense, so a typical compare() is one line, like constrain().
  • Comparing objectives that differ in direction, float step or size throws SpaceIncompatibleObjectives. DynamicCastFailed remains for spaces of the wrong type.
  • The experimental HeuristicGuidance now uses Gecode::Sense instead of its own MINIMIZE/MAXIMIZE enum.
  • The MiniModel integer, lexicographic and float optimization spaces, FlatZinc, car-sequencing and cartesian-heart implement compare(). Float spaces compare the bound their cuts use, ignoring the step. A FlatZinc satisfaction model throws SpaceNoComparison.
  • cartesian-heart has no objective; its compare() is a temporary stand-in that keeps solutions apart and will be reworked.

Tests:

  • Comparison tests cover the engines, portfolios, external incumbents, failure transport, FlatZinc and the comparison helpers.
  • Parallel BAB tests check that a weaker external bound still drops the pending solutions it excludes, and that a stored worker failure is reported first by constrain() and nogoods(). They drive the engine directly, so NoNGL is now exported for shared-library builds.
  • RBS and worker-control tests had relied on BAB enumerating equal solutions, which no longer happens. They now check strict improvement.
  • Static counters shared across tests caused races under the concurrent harness. Each test now has its own counter.

Verified locally on macOS: check and the search and FlatZinc tests pass in both the default and the experimental build, and the doc target builds.

The last commit removes .zdev/.

@zayenz zayenz left a comment •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • The one-argument CarSequencing::compare() hid the inherited Gist display overload. Added using Script::compare to restore the two-argument call.
  • Parallel PBS broadcast a solution before handling a stored worker failure. That could call into a failed asset and replace the original exception. The failure guard now runs before solution processing.

@zayenz
zayenz force-pushed the feature/comparison branch from fdd5080 to 08b56e6 Compare October 8, 2026 15:15
@zayenz
zayenz force-pushed the feature/comparison branch from 65e455e to 5aece80 Compare November 3, 2026 12:04
@zayenz zayenz changed the title Provisional: non-mutating solution comparison for Gecode 7 (breaking changes) Rank solutions with Space::compare in best-solution search (Gecode 7, ABI break) Nov 3, 2026
@zayenz zayenz changed the title Rank solutions with Space::compare in best-solution search (Gecode 7, ABI break) Rank solutions with Space::compare in best-solution search Nov 3, 2026
@zayenz
zayenz force-pushed the feature/comparison branch 2 times, most recently from 9c72d67 to 2fdfdc9 Compare November 3, 2026 16:17
@zayenz
zayenz force-pushed the feature/comparison branch from 2fdfdc9 to d7c26e2 Compare November 12, 2026 06:32
Zdev-Change-Id: Zd450758a16d392690607cf153a5160083ad5c3a5cd6812f461a16610075fc80a
…iness

Zdev-Change-Id: Z5c772894bb48c0852974820982b13020d663b002850618c6bcba63a53e9e4571
Zdev-Change-Id: Zd3bd5210f7a14908a3b4cf47a040ce93d4488f82ec178872c8b90824895e3f87
Zdev-Change-Id: Z1c007f6a8f418dab7835ce3c161f32811819842b08897f39166326aacd7afcb5
Zdev-Change-Id: Z87d7de4220478ed2e6d58d55f4648cde323acfbbb28ecb589c39215a802aeb03
Zdev-Change-Id: Z738bbca9dd306371480ab382cbd9492a52cd416d460e53b9512a5f83e57476d7
Zdev-Change-Id: Zc9f2540cb25a8b2bc131505b4314c6ea435ced462e808cb28c0b3f57b66f8f34
Arbitrate candidates and external incumbents with non-mutating comparison, invalidate superseded queued results, and deliver comparison failures safely through next().

Preserve restart support by clearing failure state during reset, with focused parallel BAB regressions.

Zdev-Change-Id: Z7926867238a6c0cfb64f226c42bf1df0d6994fedf653be929d33ef60ec652f92
Use non-mutating comparison for PBS incumbent arbitration and safely return collector or nested-asset failures after portfolio workers quiesce.

Add mixed portfolio regressions and document model-defined comparability, float ranking, migration, and the possible future best-effort incomparability policy.

Zdev-Change-Id: Ze012cd6e1d4ae7aebf1483d8aabd40f2204a99b09907e93c1a7e61d6dfc7b911
Zdev-Change-Id: Z4e4e39fd1665c0addf3e5726ab9bfb5ef3c6d49a7ca7b636d665bb672dccf5c5
Zdev-Change-Id: Z779d79cd0461e169a13d69dddf214ef27fe9499b996e28b9f44c29b66ff7cb97
Parallel BAB now discards solutions that compare as equivalent, so BAB
tests can no longer enumerate all solutions of a model without an
objective. The restart and worker-control pause tests now give BAB a
lexicographic objective and check for improving solutions ending with the
best one. The restart test model tightens its bound on the last solution
when it has an objective.

The comparison tests counted spaces and posted constraints in static
counters shared by tests that the harness runs concurrently. Give each
test its own atomic counter instead.
A* and CABS chose between solutions by their guidance bound, unlike the
other best-solution engines and the portfolios they can run in. They now
replace the incumbent only when Space::compare reports SC_BETTER, both
for their own solutions and for external incumbents, and report
incomparable and missing comparisons like BAB. Parallel workers pass
solutions that tie with the incumbent bound on to comparison. The bound
still orders and prunes open nodes.
Space::compare now says what compare returns, how it must agree with
constrain, and what the default does, at about the length of the constrain
documentation. Model-specific ranking stays with the MiniModel and FlatZinc
classes. The BAB and PBS notes and the changelog entry drop the repeated
and speculative parts.
Implementing Space::compare for an integer objective took four lines of
value comparisons, against a single rel call in constrain. compare_int
compares two values for a Sense, an optimization direction now shared by the
kernel, MiniModel, FlatZinc, and the experimental heuristic guidance, which
had its own MINIMIZE/MAXIMIZE enum.
FlatZincSpace::compare threw DynamicCastFailed for a satisfaction model,
the same error as for incompatible objectives. A model without an
objective does not support comparison, so throw SpaceNoComparison like
the default Space::compare.
The example overrides constrain to keep later solutions apart from
earlier ones, but had no compare, so parallel BAB threw
SpaceNoComparison. Report a solution as better than another exactly when
constrain would admit it. This is not an order and only keeps the
example running until it is reworked.
Parallel BAB reports a stored worker failure before constrain() compares
the new bound, and before a restart engine moves on from nogoods(). A
bound weaker than the incumbent still drops the queued solutions it
excludes. PBS reports a stored failure before broadcasting a solution to
the remaining assets, which could otherwise replace the original
exception.

The tests drive the parallel BAB engine directly, so NoNGL is exported
for shared-library builds.
A one-argument compare() hides Script's compare(const Space&,
std::ostream&), which Gist calls, so car-sequencing and cartesian-heart
bring it back into scope.
SpaceComparison becomes an enum class with Better, Equivalent, Worse and
Incomparable, and Sense uses Minimize and Maximize, matching the Gecode 7
naming of scoped enums.

Comparing solutions whose objectives differ in direction, float step or
size now throws SpaceIncompatibleObjectives instead of DynamicCastFailed
or ArgumentSizeMismatch. DynamicCastFailed remains for spaces of the wrong
type.

compare_float() and compare_lex() join compare_int(). The MiniModel float
and lexicographic spaces and FlatZinc use them.

Also declare each variable separately and clarify the FlatZinc float
comparison documentation.
@zayenz
zayenz force-pushed the feature/comparison branch from d7c26e2 to b767864 Compare November 15, 2026 08:02

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant