Implementation of hidden constraints handling by relf · Pull Request #377 · relf/EGObox

relf · 2026-01-28T15:33:11Z

This PR introduces hidden (aka unknown or crash) constraints handling
The implemention is based on hidden constraints handling described in this paper

Egor manages function evaluation crashes in two ways:

Rejection (default): where objective value and corresponding are is simply discarded
Imputation: the objective is an imputed value with penalized prediction from the objective function surrogate

# Python
egor = Egor(failsafe_strategy=egx.Egor(failsafe_strategy=egx.FailsafeStrategy.IMPUTATION)

// Rust
egor = EgorBuilder::optimize(fobj).configure(|cfg| { cfg.failsafe_strategy(FailsafeStrategy::Imputation) });

relates to #376

Copilot

Pull request overview

This PR implements hidden (aka unknown or crash) constraints handling in the EGObox optimizer, addressing issue #376. The implementation is based on the approach described in Bussemaker et al. (2024) paper. When objective function evaluations fail (return NaN), the optimizer can now either reject the failed point (default) or impute the objective value using a penalized surrogate prediction.

Changes:

Added a new FailsafeStrategy enum with two variants: Rejection (default) and Imputation
Implemented NaN detection and filtering in both initial DOE and optimization iterations
Added penalized prediction computation (pred + var for objective) to handle failed points during imputation
Extended state management to track failed points throughout optimization
Added comprehensive tests and examples demonstrating the new functionality

Reviewed changes

Copilot reviewed 20 out of 20 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
`crates/ego/src/types.rs`	Defines `FailsafeStrategy` enum for hidden constraints handling
`crates/ego/src/solver/egor_config.rs`	Adds failsafe_strategy configuration option
`crates/ego/src/solver/egor_state.rs`	Adds x_fail field to track failed evaluation points and helper methods
`crates/ego/src/solver/egor_solver.rs`	Filters NaN values from initial DOE and initializes failed points tracking
`crates/ego/src/solver/solver_impl.rs`	Implements penalized prediction imputation logic for failed points
`crates/ego/src/solver/solver_computations.rs`	Adds compute_penalized_point method for crash handling
`crates/ego/src/solver/trego.rs`	Integrates failsafe strategy into TREGO algorithm
`crates/ego/src/utils/misc.rs`	Implements filter_nans, usable_data, and updated update_data functions
`crates/ego/src/utils/portfolio.rs`	Updates portfolio signature to include penalized values
`crates/ego/src/solver/solver_infill_optim.rs`	Adds constructor for InfillOptProblem
`crates/ego/src/egor.rs`	Adds comprehensive tests for failsafe strategies
`crates/ego/src/lib.rs`	Updates documentation with reference to Bussemaker2024 paper
`python/src/types.rs`	Defines Python FailsafeStrategy enum
`python/src/lib.rs`	Registers FailsafeStrategy class in Python module
`python/src/egor.rs`	Adds failsafe_strategy parameter to Egor Python class
`python/egobox/egobox.pyi`	Updates type stubs with FailsafeStrategy enum and parameter
`python/egobox/tests/test_egor.py`	Adds test for constrained Branin with NaN handling, adjusts tolerances
`python/egobox/examples/branin_nans.py`	New example demonstrating hidden constraints via NaN returns
`python/egobox/examples/branin_fcstr.py`	New example showing standard constraint handling for comparison
`README.md`	Adds reference to Bussemaker2024 paper

Copilot · 2026-01-30T11:08:48Z

crates/ego/src/solver/solver_impl.rs

+
+                    // update_data(
+                    //     &mut x_dat, &mut y_dat, &mut c_dat, x_new, &y_new, &c_new, None,
+                    // );


There is commented-out code that should be removed. This dead code appears to be leftover from development and serves no purpose in the final implementation.

Suggested change

// update_data(

// &mut x_dat, &mut y_dat, &mut c_dat, x_new, &y_new, &c_new, None,

// );

Copilot · 2026-01-30T11:08:48Z

python/egobox/tests/test_egor.py

+                    return branin(np.atleast_2d(xi))[0, 0]
+
+            res = np.apply_along_axis(branin_constraint_nans, 1, x).reshape(-1, 1)
+            print(x, res)


Debug print statement should be removed from test code. This print statement at line 393 appears to be leftover from debugging and should be removed for cleaner test output.

Suggested change

print(x, res)

Copilot · 2026-01-30T11:08:49Z

crates/ego/src/lib.rs

 //! * Theta bounds are implemented as in \[[Appriou2023](#Appriou2023)\]
-//! * Logirithm of Expected Improvement is implemented as in \[[Ament2025](#Ament2025)\]
+//! * Logarithm of Expected Improvement is implemented as in \[[Ament2025](#Ament2025)\]
+//! * Hidden constraints handling are implemented as in \[[Bussemaker2024](#Bussemaker2024)\]


Grammatical error in documentation: "are" should be "is". The phrase should read "Hidden constraints handling is implemented as in" rather than "are implemented as in" since "handling" is singular.

Suggested change

//! * Hidden constraints handling are implemented as in \[[Bussemaker2024](#Bussemaker2024)\]

//! * Hidden constraints handling is implemented as in \[[Bussemaker2024](#Bussemaker2024)\]

Copilot · 2026-01-30T11:08:49Z

crates/ego/src/solver/solver_computations.rs

+    /// * First element of the tuple is the objective function values
+    /// * Second element of the tuple is the penalized objective function values
+    ///   used in case of true evaluation crash
+    ///   This a the penalized prediction (pred + var) for objective


Typo in comment: "This a the" should be "This is the". The comment should read "This is the penalized prediction" instead of "This a the penalized prediction".

Suggested change

/// This a the penalized prediction (pred + var) for objective

/// This is the penalized prediction (pred + var) for objective

Copilot · 2026-01-30T11:08:49Z

python/egobox/tests/test_egor.py

+    unittest.main(
+        defaultTest=["TestEgor.test_constrained_branin_with_nans"], exit=False
+    )


The defaultTest parameter should be removed or commented out before merging to main. The change on lines 413-415 sets a specific test as the default test to run, which is useful during development but should not be committed to the main branch as it prevents running the full test suite by default.

Suggested change

unittest.main(

defaultTest=["TestEgor.test_constrained_branin_with_nans"], exit=False

)

unittest.main(exit=False)

Copilot · 2026-01-30T11:08:50Z

python/egobox/tests/test_egor.py

+        self.assertAlmostEqual(18.935, res.x_opt[0], delta=5e-2)

    def test_xsinx_with_reclustering(self):
        egor = egx.Egor([[0.0, 25.0]], seed=42, gp_config=egx.GpConfig(n_clusters=0))
        res = egor.minimize(xsinx, max_iters=20)
        print(f"Optimization f={res.y_opt} at {res.x_opt}")
        self.assertAlmostEqual(-15.125, res.y_opt[0], delta=1e-3)
-        self.assertAlmostEqual(18.935, res.x_opt[0], delta=1e-2)
+        self.assertAlmostEqual(18.935, res.x_opt[0], delta=5e-2)


Test tolerance relaxation should be justified or reverted. The delta tolerance was increased from 1e-2 to 5e-2 for x_opt assertions in the xsinx tests (lines 169 and 176). This relaxation of test tolerances should either be justified with a comment explaining why it's necessary (e.g., if the hidden constraints handling affects convergence behavior) or reverted if it's an unintended change. Relaxing test tolerances without justification can mask actual regressions.

Copilot · 2026-01-30T11:08:50Z

python/egobox/examples/branin_nans.py

+            print(f"Constraint violated at {xi}, returning NaN")
+            return np.nan
+        else:
+            print(branin_forrester(np.atleast_2d(xi)))
+            return branin_forrester(np.atleast_2d(xi))


Debug print statements should be removed from the example. The print statements at lines 54 and 57 within the constraint evaluation function appear to be debug traces and should be removed for cleaner output in the example.

Suggested change

print(f"Constraint violated at {xi}, returning NaN")

return np.nan

else:

print(branin_forrester(np.atleast_2d(xi)))

return branin_forrester(np.atleast_2d(xi))

return np.nan

else:

value = branin_forrester(np.atleast_2d(xi))

return value

Copilot · 2026-01-30T11:08:50Z

crates/ego/src/solver/solver_impl.rs

+                            let y_pred = self
+                                .compute_penalized_point(&xfail, models[0].as_ref(), &models[1..])
+                                .unwrap();


Use of unwrap() without error handling could cause panic. The call to compute_penalized_point() at line 679 uses unwrap(), which could panic if the prediction fails. Since this is in a critical path (handling failed points during optimization), consider propagating the error or handling it more gracefully. The existing pattern in the codebase (line 808-810) shows that Result can be returned and handled properly.

Copilot · 2026-01-30T11:08:50Z

python/egobox/tests/test_egor.py

    def test_g24_with_fcstrs(self):
        n_doe = 5
-        max_iters = 20
+        max_iters = 5


Reduction in test iterations reduces test coverage. The max_iters was reduced from 20 to 5 for the g24_with_fcstrs test (line 324). This significant reduction in the number of optimization iterations may reduce the test's ability to catch convergence issues or regressions. If this change was made to speed up tests, consider whether the reduced coverage is acceptable or if this test should be split into a quick smoke test and a more thorough integration test.

relf added 17 commits January 24, 2026 13:34

Store objective computation failures (NaN values)

4f71975

Linting

c6e321e

Add failure handling test

4de6b71

Add missing must_use

49d39f0

Seed test to avoid occasional failure

9977ba7

Add best_index method

9081f84

Remove infill_data from select_next_point return

e36ae4f

Implement PredictionImputation failsafe strategy

79aba88

Impute nan found in the initial doe wrt failsafe strategy

c4616af

Add Python API for fail safe strategy

c336bc7

Add documentation

da78446

Fix tests

71f5f4c

Linting

dbf0d47

Fix doc syntax

efc6309

Fix method name

b70e354

Try to debug CI test (it works locally)

285788a

Skip test that fails on CI

9733b44

relf marked this pull request as ready for review January 30, 2026 09:12

relf requested a review from Copilot January 30, 2026 11:03

Copilot started reviewing on behalf of relf January 30, 2026 11:03 View session

Copilot AI reviewed Jan 30, 2026

View reviewed changes

Cleanup

123f0a2

relf merged commit f37285e into master Jan 30, 2026
19 checks passed

relf deleted the handle-nan-obj branch January 30, 2026 14:15

This was referenced Feb 2, 2026

Add probability of viability constraint as failsafe strategy #378

Merged

Hidden constraints not handled #376

Closed

relf mentioned this pull request Feb 19, 2026

Roadmap #90

Open

17 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implementation of hidden constraints handling#377

Implementation of hidden constraints handling#377
relf merged 18 commits intomasterfrom
handle-nan-obj

relf commented Jan 28, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Copilot AI Jan 30, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	//! * Hidden constraints handling are implemented as in \[[Bussemaker2024](#Bussemaker2024)\]
	//! * Hidden constraints handling is implemented as in \[[Bussemaker2024](#Bussemaker2024)\]

	/// This a the penalized prediction (pred + var) for objective
	/// This is the penalized prediction (pred + var) for objective

Conversation

relf commented Jan 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

relf commented Jan 28, 2026 •

edited

Loading