Minimal Dependencies

Gradient-Free-Optimizers is designed to have a small dependency footprint. The core library requires only packages that are already present in virtually every Python data science environment.

Core Dependencies

The following packages are required:

  • NumPy – array operations and search space representation

  • pandas – results storage via opt.search_data DataFrame

  • SciPy – used internally by some optimizers (e.g. distance calculations)

  • tqdm – progress bar display during search()

That’s it. No heavy frameworks, no compiled C extensions beyond what NumPy and SciPy already provide, no external optimization backends.

Optional Dependencies

For surrogate-model-based optimizers (BayesianOptimizer, ForestOptimizer, TreeStructuredParzenEstimators), you can optionally install scikit-learn to use sklearn-based surrogate models:

pip install gradient-free-optimizers[sklearn]

This adds:

  • scikit-learn – provides alternative GP, Random Forest, and KDE implementations for surrogate models

However, sklearn is not required. GFO ships its own Gaussian Process, Random Forest, and KDE implementations, so all 22 optimizers work out of the box without scikit-learn installed.

Why This Matters

Fast installation. pip install gradient-free-optimizers completes in seconds. There are no large compiled dependencies to build or download beyond the standard scientific Python stack.

Small footprint. The flat dependency tree means no transitive dependency surprises. You know exactly what gets installed.

Works everywhere. The minimal requirements mean GFO runs reliably in constrained environments:

  • Docker containers with slim base images

  • CI/CD pipelines where install time matters

  • Serverless functions with package size limits

  • Air-gapped environments with limited package availability

No conflicts. NumPy, pandas, SciPy, and tqdm are among the most widely used Python packages. Adding GFO to an existing project is unlikely to introduce version conflicts.

Built-in Surrogate Models

A key design decision: GFO provides its own implementations of the surrogate models used by sequential model-based optimizers. This means BayesianOptimizer works immediately after installation, without scikit-learn:

from gradient_free_optimizers import BayesianOptimizer
import numpy as np

search_space = {
    "x": np.linspace(-10, 10, 100),
    "y": np.linspace(-10, 10, 100),
}

def objective(para):
    return -(para["x"] ** 2 + para["y"] ** 2)

# Works without sklearn -- uses GFO's built-in GP
opt = BayesianOptimizer(search_space)
opt.search(objective, n_iter=50)
print(opt.best_para)

If sklearn is available, you can pass an sklearn estimator via the gpr parameter for BayesianOptimizer, or configure the tree-based surrogates via tree_regressor and tree_para in ForestOptimizer.