Add pluggable Threshold-Elites and Dominated Novelty Search (DNS) database variants alongside MAP-Elites

[fangyihao]

Summary

This change makes the program database implementation configurable at runtime and introduces two new database strategies—DNSProgramDatabase and ThresholdElitesProgramDatabase—while tightening several multi-island correctness edge cases (copy embedding, infinite-score handling). It also adds dedicated unit tests for both new database variants.

What changed

openevolve/config.py

• Extended DatabaseConfig with:

  • variant: str = "DNSProgramDatabase" to select the database implementation by class name.

  • distance_threshold: float = 1.5 (used by the threshold-based elite filtering).

openevolve/controller.py

• Switched database construction from a fixed ProgramDatabase(...) to dynamic instantiation:

  • Reads config.database.variant (defaults to "ProgramDatabase" if missing).

  • Uses importlib to load openevolve.database. and instantiate it.

• This enables selecting different database behaviors via config without changing code.

openevolve/database.py

General robustness fixes

• Import copy and math.

• _update_best_program: avoid comparing / logging combined_score when either side is ±inf (prevents misleading “best program changed” logs and edge-case issues).

• When seeding/initializing empty islands via copying the best program, the copied program now deep-copies embedding to avoid shared references across programs/islands.

• get_island_stats: filters out ±inf scores before computing best_score / avg_score for cleaner island stats.

New database variants

• ThresholdElitesProgramDatabase

  • Computes feature embedding + fitness-derived combined_score.

  • Runs novelty gating via existing _is_novel.

  • Applies a distance-threshold competition inside an island:

    • For programs within distance_threshold, the lower-scoring one gets combined_score = -inf (effectively removed from competition).

• DNSProgramDatabase

  • Implements Dominated Novelty Search-style selection:

    • For each program, considers the set of dominating programs (higher combined_score).

    • Computes dominated novelty score as mean distance to those dominating programs (or +inf if none dominate).

    • Uses the median dominated novelty score as a cutoff; programs below the median get combined_score = -inf.

  • Keeps island assignment consistent and respects novelty gating before DNS filtering.

Tests

Added two new test suites:

• tests/test_dns_database.py

• tests/test_threshold_elites_database.py

These validate:

• Multi-island initialization (correct island counts, per-island best tracking).

• Targeted island insertion + metadata["island"] correctness.

• Migration behavior: no migrant suffix IDs and migrants don’t get re-migrated.

• Empty-island sampling creates copies (not shared references) and passes validation.

• No program is ever assigned to multiple islands.

• Migration validation remains stable across multiple migration cycles (prevents exponential “migrant” growth).

How to use

Set in config:

• database.variant = "DNSProgramDatabase" (default in this diff)

• or database.variant = "ThresholdElitesProgramDatabase"

• optionally tune database.distance_threshold for the Threshold-Elites behavior.

Testing

• Added unit tests for both variants (DNS + Threshold-Elites).

• Existing database test base class is reused to ensure behavior parity where intended.