Nevergrad — Gradient-Free Optimization by Meta

Introduction

Nevergrad is a Python library from Meta AI Research that provides gradient-free optimization algorithms under a common interface. It is designed for problems where gradients are unavailable or unreliable, such as hyperparameter tuning, reinforcement learning reward shaping, and simulation-based optimization.

What Nevergrad Does

• Optimizes black-box functions without requiring gradients
• Provides 30+ optimization algorithms under a unified API
• Supports continuous, discrete, and mixed search spaces
• Offers built-in benchmarks for comparing optimizer performance
• Handles noisy objective functions with appropriate averaging

Architecture Overview

Nevergrad defines a Parametrization system that describes the search space (scalars, arrays, choices, log-scales). An Optimizer wraps a specific algorithm (CMA-ES, differential evolution, PSO, etc.) and generates candidates via ask/tell. The ask method proposes parameter values; tell reports back the loss. This separation allows asynchronous and parallel evaluation.

Self-Hosting & Configuration

• Install via pip with no heavy dependencies
• Define search spaces using ng.p.Scalar, ng.p.Array, or ng.p.Choice
• Select an optimizer or use NGOpt for automatic algorithm selection
• Set budget to control the total number of objective evaluations
• Use num_workers > 1 to evaluate candidates in parallel batches

Key Features

• NGOpt meta-optimizer automatically picks the best algorithm for your problem
• Expressive parametrization supporting constraints and transformations
• Ask/tell interface enables asynchronous and distributed evaluation
• Built-in benchmarking suite for rigorous optimizer comparison
• Supports multi-objective optimization via Pareto front tracking

Comparison with Similar Tools

• Optuna — Bayesian optimization with pruning for ML; Nevergrad covers broader optimization use cases beyond ML
• Hyperopt — TPE-based search for hyperparameters; Nevergrad offers more diverse algorithms
• scipy.optimize — classical numerical optimization; Nevergrad handles noisy, non-differentiable objectives
• Ray Tune — orchestrates trials at scale; Nevergrad focuses on the optimization algorithms themselves

FAQ

Q: What is NGOpt? A: NGOpt is a meta-optimizer that selects the best algorithm based on your problem characteristics (budget, dimensionality, noise level).

Q: Can Nevergrad optimize discrete variables? A: Yes. Use ng.p.Choice for categorical variables and ng.p.TransitionChoice for ordered discrete parameters.

Q: How do I run evaluations in parallel? A: Set num_workers and call ask() multiple times before calling tell() with results as they arrive.

Q: Does Nevergrad support constraints? A: Yes. Apply register_cheap_constraint to penalize infeasible regions of the search space.

Sources

• https://github.com/facebookresearch/nevergrad
• https://facebookresearch.github.io/nevergrad/