Lightweight Covariance Matrix Adaptation Evolution Strategy (CMA-ES)

0.13.0 · active · verified Sat Apr 11

cmaes is a lightweight Python library providing an implementation of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) for numerical optimization. It offers a simple "ask-and-tell" interface and has expanded to include various advanced CMA-ES variants such as CatCMA, CMAwM, and COMO-CatCMAwM, catering to mixed-variable and multi-objective optimization problems. The library is actively maintained, currently at version 0.13.0, with frequent updates introducing new algorithms and features.

Warnings

breaking Python 3.6 support was dropped in `v0.9.1`.
Fix: Upgrade Python environment to 3.8 or later. The current recommended version is >=3.9.
gotcha The library offers numerous specialized optimizers (e.g., `CatCMAwM`, `COMOCatCMAwM`, `CMAwM`) for specific problem types beyond continuous, single-objective optimization.
Fix: Carefully select and import the appropriate optimizer class (e.g., `from cmaes import CatCMAwM`) based on your problem's characteristics (e.g., mixed-variables, multi-objective).
gotcha The core `CMA` class is designed for continuous optimization problems. Using it directly for problems involving integer, categorical, or mixed variables, or for multi-objective optimization, may lead to suboptimal performance or incorrect results.
Fix: For mixed-variable problems, consider `CatCMAwM`. For multi-objective problems, `COMOCatCMAwM` is available.

Install

pip install cmaes PyPI
conda install -c conda-forge cmaes Conda-forge

Imports

CMA
```
from cmaes import CMA
```
The primary CMA-ES optimizer is directly available from the top-level package.
CatCMAwM
```
from cmaes import CatCMAwM
```
For mixed-variable (continuous, integer, categorical) optimization problems.
COMOCatCMAwM
```
from cmaes import COMOCatCMAwM
```
For multi-objective mixed-variable optimization problems, typically 2 objectives.

Quickstart

This example demonstrates the basic "ask-and-tell" interface of the `cmaes` library using the standard `CMA` optimizer to minimize a simple quadratic function. The optimizer generates candidate solutions (`ask`), these are evaluated by the objective function, and then the results (`tell`) are used to update the search distribution for the next generation.

import numpy as np
from cmaes import CMA

def quadratic(x1, x2):
    # Objective function to minimize
    return (x1 - 3) ** 2 + (10 * (x2 + 2)) ** 2

if __name__ == "__main__":
    # Initialize CMA optimizer with an initial mean and standard deviation
    optimizer = CMA(mean=np.zeros(2), sigma=1.3)

    print("Starting CMA-ES optimization...")
    for generation in range(50):
        solutions = []
        # Ask for new candidate solutions
        for _ in range(optimizer.population_size):
            x = optimizer.ask()
            value = quadratic(x[0], x[1])
            solutions.append((x, value))
        
        # Print the best value found in the current generation
        # (Note: CMA-ES updates distribution based on all solutions, not just the best)
        print(f"Generation #{generation+1}: Best value = {min(s[1] for s in solutions):.4f}")
        
        # Tell the optimizer the evaluated solutions and their values
        optimizer.tell(solutions)

    print(f"\nOptimization finished. Final mean of the search distribution: {optimizer.mean}")

view raw JSON →