Threshold-Elites Island-based Program Selection

[fangyihao]

Net effect (what changes for users)

• You can now choose between map-elites and threshold-elites archive insertion using DatabaseConfig.variant.

  • In threshold-elites, programs are:

  • filtered by novelty,

  • grouped into islands (with parent island inheritance),

• and locally compared by embedding distance with a winner-takes-all rule inside a threshold radius.

• The attention optimization evaluator becomes more portable (pathing) and its output schema shifts toward a combined_score field.

openevolve/database.py

This is the main functional change: introduces a “threshold-elites” variant alongside map-elites, and refactors add() accordingly.

Variant-based dispatch in add(...)

• Database.add(...) now checks config.variant:

  • "map-elites" → add_map_elites_program(...)

  • "threshold-elites" → add_threshold_elites_program(...)

  • otherwise raises NotImplementedError.

New add_threshold_elites_program(...) behavior

When adding a program:

• Stores program + iteration bookkeeping

  • Updates program.iteration_found and self.last_iteration.

• Computes embedding + an elite score

  • Calls _calculate_feature_coords(program) and stores in program.embedding.

  • Computes a fitness-based score = get_fitness_score(...).

  • Stores program.metrics["elite_score"] = score.

• Determines island placement

  • If target_island is not given and the program has a parent, it inherits the parent’s island from parent.metadata["island"] to keep island isolation.

  • Else falls back to self.current_island.

• Novelty gating happens before insertion

  • If _is_novel(...) fails, the program is not added to the island.

• Threshold-elites competition (within an island)

  • For existing programs in the island, computes Euclidean distance between embeddings.

  • If distance < elite_threshold, it keeps the higher elite_score and demotes the other by setting its elite_score to -inf (effectively marking it as dominated).

• Finalization

  • Adds the program to the island, sets program.metadata["island"], updates archive/best tracking, enforces population limit, and optionally saves to disk.

Exploration parent sampling now preserves embeddings

• In two places inside _sample_exploration_parent(...), when cloning the best program, it now does:

  • embedding=copy.deepcopy(best_program.embedding)

• This prevents the clone from sharing the same embedding object reference.

Misc

• copy, math, asdict/fields imports added (supporting new logic).

openevolve/config.py

• Extends DatabaseConfig with parameters to support a new archive strategy:

  • variant: str = "threshold-elites" (strategy selector)

  • elite_threshold: int = 3 (distance threshold used by the new strategy)

examples/attention_optimization/evaluator.py

• Makes MLIR path resolution robust:

  • Previously: Path("mlir/self_attn_with_consts_linalg_dialect.mlir") (relative to current working directory).

  • Now: builds an absolute path relative to the evaluator file’s directory using file. This prevents failures when running from a different working directory.

• Renames/standardizes a scoring field:

  • In estimate_performance_from_ir(...), the returned dict changes from using optimization_score to combined_score.

  • In evaluate(...), result packaging similarly switches from optimization_score → combined_score.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ fangyihao
❌ Yihao Fang

---

Yihao Fang seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[fangyihao]

Converting this PR to Draft temporarily because the one of the commits isn't currently attributed to the correct GitHub account (author email not linked). Once the author email is updated/verified and the commits show proper attribution, we’ll move it back to Ready for review.

[fangyihao]

Closing this PR for now because the one of the commits is not attributed to the correct GitHub account (author linking/email mismatch), which breaks ownership/traceability for review and auditing.