Skip to content

feat(compat): adapt local AlphaEvolve evaluators for OpenEvolve - #519

Merged
codelion merged 1 commit into
algorithmicsuperintelligence:mainfrom
Yushu-Li:feat/alphaevolve-evaluator-interop
Oct 19, 2026
Merged

codelion merged 1 commit into
algorithmicsuperintelligence:mainfrom
Yushu-Li:feat/alphaevolve-evaluator-interop

Conversation

@Yushu-Li

Copy link
Copy Markdown
Contributor

OpenEvolve evaluators receive a file path, while Google's local AlphaEvolve examples expect a candidate object and return nested scores/insights. Add an explicit, dependency-free adapt_evaluator helper under openevolve.compat so those evaluators can be used unchanged through the existing callable API.

Partially addresses #468: this is the A1 evaluator-contract slice and a local walkthrough. It does not close the issue or provide SDK emulation, config translation, export, or cloud execution.

  • Preserve all metrics and insights; require an explicit objective for combined_score. Map configurable failure sentinels/null/non-finite scores to failure metrics without accidentally promoting them above valid negative objectives.
  • Add an evaluate-only runner for the upstream circle-packing, TSP and signal-processing examples, with opt-in evolution using the user's OE config. Existing evaluator/config defaults are untouched; no Google dependency is added.

Validation:

  • Full python -m unittest discover -s tests -q: 687 tests passed, including 15 new adapter tests and a real controller/worker round-trip with a deterministic provider.
  • Black, isort and git diff --check pass.
  • Against Google-Cloud-AI/alphaevolve-on-googlecloud at 674dd5e558e917320c251c085e47fae33e1bbcff: all three unchanged seed/evaluator pairs evaluated locally and completed one OE generation through real worker processes using scripted proposals. Original input files remained byte-identical. Network connections were blocked by the validation harness. This verifies interoperability, not learned optimization quality; no paid LLM or real Google service was used.

Limitations are documented: synchronous single-file local evaluators only, no sandbox, and exceptions/timeouts retain OE's existing generic error handling (which can return zero fitness).

AI disclosure: Codex assisted with implementation, tests and documentation for a user-selected existing public issue found during contribution research. All validation reported above was executed locally.

@CLAassistant

CLAassistant commented Oct 19, 2026 •

Copy link
Copy Markdown

CLA assistant check
All committers have signed the CLA.

@codelion
codelion merged commit 34427ba into algorithmicsuperintelligence:main Oct 19, 2026
4 checks passed
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.

3 participants