Research software
HOLA
A lightweight asynchronous system for multi-objective black-box and hyperparameter optimization, with a Rust engine, Python interface, and a common ask-and-tell protocol for local or distributed studies.
Current state
The public HOLA codebase provides a Rust-backed optimization engine, Python API, REST and command-line interfaces, distributed workers, checkpointing, and a browser dashboard. The default optimization strategy uses a kind of cross-entropy method. After initial Sobol exploration, its proxy distribution is a GMM fit on the top-performing samples so far, with periodic global Sobol exploration continuing afterward. Random and pure Sobol strategies are also available.
Future plans
I still need to update the arXiv paper so that it documents the current HOLA implementation and results. The 2022 version still documents the old Python library and does not describe objective grouping.
Details
- Phase
- Public 1.0 release candidate · arXiv revision pending
- Domain
- Black-box optimization · hyperparameter optimization · distributed systems
- My role
- Developer
Artifacts
- Repository repository
- 2022 arXiv manuscript preprint