CyberAgentAILab/cmaes

What it solves

It provides a simple and practical Python implementation of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) and its various specialized variants. It is designed to solve complex optimization problems, including those with mixed-variable spaces (continuous, integer, and categorical) and those requiring safety constraints or multi-objective goals.

How it works

The library uses an "ask-and-tell" interface where the optimizer suggests candidate solutions (ask) and the user provides the resulting objective values (tell). It implements the standard CMA-ES algorithm and several advanced extensions:

• CatCMAwM: Handles mixed-variable optimization (continuous, integer, and categorical) in a unified framework.
• COMO-CatCMAwM: Coordinates multiple optimizers to perform multi-objective mixed-variable optimization.
• Safe CMA: Uses safety functions and safe seeds to optimize while avoiding solutions that exceed safety thresholds.
• LRA-CMA: Automatically adjusts the learning rate to handle multimodal or noisy problems without manual hyperparameter tuning.
• WS-CMA: Allows transferring prior knowledge from similar tasks to speed up optimization when evaluation budgets are limited.

Who it’s for

• Researchers and developers working on black-box optimization.
• Machine learning engineers performing hyperparameter optimization (it integrates with Optuna).
• Engineers dealing with mixed-variable search spaces or safety-critical optimization tasks.

Highlights

• Mixed-Variable Support: Specialized variants for continuous, integer, and categorical variables.
• Safety-Aware: Includes a Safe CMA-ES variant for constrained optimization.
• Multi-Objective: Supports two-objective mixed-variable optimization via COMO-CatCMAwM.
• Auto-Tuning: LRA-CMA reduces the need for expensive hyperparameter configuration.
• Integration: Compatible with the Optuna framework for automatic hyperparameter optimization.