Nevergrad

Pauline Bennet, Carola Doerr, A. Moreau, Jérémy Rapin, Fabien Teytaud, Olivier TeytaudACM SIGEVOlution 2021 · DOI

Nevergrad Logo

Abstract

Nevergrad is an open source platform for black-box optimization. Join the user group! And if you like Nevergrad, please support us by adding a star on GitHub (https://github.com/facebookresearch/nevergrad).

The Challenge of Black-Box Optimization

In many engineering and scientific domains—ranging from structural design to hyperparameter tuning for neural networks—we often encounter problems where the objective function is a "black box." We can evaluate the quality of a solution, but we lack access to gradients or a closed-form mathematical expression. This is the realm of black-box optimization.

Historically, this field has struggled with two primary issues. First, there is an algorithm overload: with hundreds of techniques available (e.g., evolutionary strategies, Bayesian optimization), practitioners often resort to trial and error to find the right tool for their specific constraints. Second, benchmarking fragmentation makes it difficult to compare new methods fairly, as researchers frequently use inconsistent implementations and evaluation metrics.

A Unified Optimization Framework

Nevergrad addresses these hurdles by providing a comprehensive, open-source library that bridges the gap between theoretical research and practical engineering. It serves as both a repository of state-of-the-art solvers and a robust benchmarking laboratory.

Optimization sequence

At its core, Nevergrad implements a diverse portfolio of algorithms, including Covariance-Matrix-Adaptation (CMA), Particle Swarm Optimization (PSO), and Differential Evolution (DE). For multi-objective tasks, it incorporates standard tools like NSGA-II. Beyond standard implementations, the platform introduces specialized techniques such as TBPSA for noisy environments and quasi-opposite sampling to enhance one-shot optimization in high-dimensional spaces.

Intelligent Algorithm Selection with NGOpt

One of the platform's most innovative features is NGOpt, an "optimization wizard." Rather than requiring users to be experts in optimization theory, NGOpt uses high-level problem metadata—such as the available budget, the dimensionality of the search space, and the types of variables involved—to automatically select, chain, or parallelize the most effective solvers. This "user-friendly" layer ensures that the best-performing tools are accessible to engineers without extensive tuning.

Benchmarking and Real-World Impact

Nevergrad provides a unified interface to several benchmark suites, including its own YABBOB and LSGO, as well as external integrations like MuJoCo for robotics. A unique aspect of the platform is its commitment to continuous benchmarking; rankings are periodically updated to reflect the current state-of-the-art as codebases evolve.

The practical utility of Nevergrad is evidenced by its adoption in high-stakes projects:

  • Stockfish Tuning: It was used to optimize the neural value function weights for Stockfish, helping the engine maintain its position as a top-tier chess AI.
  • Photonics: The platform enabled the design of advanced antireflective coatings for silicium, significantly outperforming traditional designs.
  • Subjective Optimization: By leveraging evolutionary algorithms, Nevergrad has been used to evolve GAN outputs based on human preference, demonstrating its versatility in tasks where the objective is subjective taste.

Future Outlook

While Nevergrad has already seen significant research impact—cited in over 70 Arxiv papers—it remains an evolving project. Future development will focus on improving complex constraint management and expanding the library of real-world application benchmarks. By continuing to refine the logic behind automatic algorithm selection, Nevergrad aims to remain the go-to platform for robust, gradient-free optimization.

Summary updated

Abstract

Nevergrad is an open source platform for black-box optimization. Join the user group! And if you like Nevergrad, please support us by adding a star on GitHub (https://github.com/facebookresearch/nevergrad).
BibTeX
@inproceedings{bennet2021nevergrad,
  author = {Pauline Bennet and Carola Doerr and A. Moreau and Jérémy Rapin and Fabien Teytaud and Olivier Teytaud},
  title = {{Nevergrad}},
  booktitle = {ACM SIGEVOlution},
  year = {2021},
  doi = {10.1145/3460310.3460312},
}
Figure