Support vector-valued combined_score for multi-objective optimization

[yuxuan-z19]

This PR introduces support for vector-valued combined_score (e.g., tuple[float, ...]) and generalizes its handling across the OpenEvolve pipeline.

Related issues:

• Refs [RFC] Enhancing combined_score to support multi-objective optimization via Lexicographical ordering #453
• Resolves Use multiple metrics in more of a Pareto optimization way? #110
• Resolves Automatic stopping when a certain metric reaches a certain value? #109

---

Changes

Fitness Representation

• combined_score now supports: float | tuple | list
• Removed implicit casting to float

Comparison Semantics

• All comparisons and deltas are handled via NumPy:
  • np.subtract
  • np.allclose
• Enables consistent element-wise behavior for non-scalar scores

Early Stopping

• Thresholds are broadcast to match vector-valued scores
• Event-based stopping uses np.allclose() for convergence checks

Logging and Formatting

• Introduced format_score() for unified scalar/vector formatting
• Removed assumptions of scalar formatting (%.4f, :.4f)

Aggregation / Statistics

• Island statistics (e.g., mean) are computed element-wise for tuple/list scores

Pipeline Integration

• Changes are applied across:
  • evolution loop
  • logging
  • prompt rendering

Testing

• Added/updated unit tests covering tuple/list combined_score
• All existing tests pass

---

Backward Compatibility

• Scalar combined_score remains fully supported
• Existing workflows continue to function without modification

Comparison logic is extended to support vector semantics when applicable.

---

Fallback Behavior

When combined_score is absent, the existing fallback is preserved: safe_numeric_average(p.metrics). This remains scalar (no implicit broadcasting).

Broadcasting this value to match vector-valued comparisons was considered, but not included to avoid introducing implicit behavior. This can be revisited if stricter consistency is preferred.

---

Summary

This PR extends combined_score from a scalar assumption to a general vector-compatible representation, with consistent behavior across evaluation, comparison, logging, and aggregation.

Regards,
@yuxuan-z19 (/w Baidu FM-Agent @baidubce)

[Copilot]

Pull request overview

Adds first-class support for vector-valued combined_score (e.g., tuple[float, ...] / list[float]) to enable multi-objective optimization semantics across scoring, logging, early stopping, and island statistics.

Changes:

• Generalize combined_score handling (fitness extraction, formatting, and prompt rendering) to support scalar and vector scores.
• Update comparison/delta and aggregation paths to use NumPy operations for non-scalar scores.
• Add unit tests and a MoE load-balancing example workload demonstrating tuple-based scoring.

Reviewed changes

Copilot reviewed 21 out of 21 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
tests/test_tuple_score.py	Adds tests validating get_fitness_score() behavior for float/list/tuple combined_score.
tests/test_database.py	Adds tests for island stats aggregation with float/list/tuple scores; minor formatting cleanups.
openevolve/utils/metrics_utils.py	Returns list/tuple combined_score as-is; adds broadcast_value() helper.
openevolve/utils/format_utils.py	Introduces format_score() and uses NumPy subtraction for improvement formatting.
openevolve/prompts/defaults/fragments.json	Removes scalar-only :.4f formatting placeholders for fitness fragments.
openevolve/prompt/sampler.py	Uses format_score() in prompt rendering; adds NumPy allclose for “stable” fitness; broadcasts thresholds for vector scores.
openevolve/process_parallel.py	Updates early-stopping logging/threshold handling to support vector-valued scores.
openevolve/database.py	Uses format_score() for logging; uses np.subtract for score diffs; computes island averages element-wise via NumPy.
openevolve/api.py	Widens EvolutionResult.best_score type and updates __repr__ formatting via format_score().
examples/moe_lb/run.sh	Adds a runnable script for the MoE load-balancing example.
examples/moe_lb/requirements.txt	Adds dependencies for the MoE load-balancing example.
examples/moe_lb/initial_program.py	Provides a baseline implementation for the MoE load-balancing task.
examples/moe_lb/evaluator.py	Adds an evaluator emitting tuple combined_score for multi-objective scoring.
examples/moe_lb/config.yaml	Adds an example config emphasizing lexicographic multi-objective scoring.
examples/moe_lb/archives/task_config.yaml.bak	Stores archived upstream task config reference.
examples/moe_lb/archives/prompt_multi_obj.md	Archived prompt text for multi-objective variant.
examples/moe_lb/archives/prompt_lp.md	Archived prompt text for LP-focused variant.
examples/moe_lb/archives/prompt_common.md	Archived prompt text for common/heuristic variant.
examples/moe_lb/archives/loongflow_best_lp.py	Archived “best” LP solution reference implementation.
examples/moe_lb/archives/loongflow_best_common.py	Archived “best” heuristic solution reference implementation.
.pre-commit-config.yaml	Updates isort and black hook revisions.

[Copilot]

For vector-valued score, the score >= HIGH_SCORE / MED_SCORE / LOW_SCORE checks rely on Python sequence comparisons, which are lexicographic rather than element-wise. This can misclassify programs (e.g., a high first objective but very low secondary objectives still counts as “high performer”). If the intent is element-wise thresholds, use NumPy comparisons (e.g., np.all(np.greater_equal(score, HIGH_SCORE)) after coercing to arrays); if the intent is lexicographic ordering, consider comparing only the primary objective explicitly to make the rule unambiguous.

[Copilot]

convergence_threshold is defined only inside the early_stopping_patience > 0 branch, but it’s referenced in the event-based early stopping branch. If early_stopping_patience <= 0, this will raise an UnboundLocalError at runtime. Define/broadcast convergence_threshold before the patience check (or reference self.config.convergence_threshold directly) so both modes work.