Nevergrad is a gradient-free optimization library and hyperparameter optimization framework designed to find the minimum of objective functions without using derivatives. It serves as an asynchronous optimization engine that decouples parameter suggestions from result reporting to support parallel function evaluations.
The project specializes in multi-objective optimization to identify Pareto fronts for competing goals and provides a suite for benchmarking the performance and convergence of different optimization routines. It supports black-box system optimization, enabling the tuning of external scripts or non-native code by injecting parameter values into source files.
The library handles a wide variety of search spaces, including continuous, discrete, and categorical variables, and can optimize noisy or ill-conditioned functions. Its capability surface includes distributed parameter search, the ability to chain multiple algorithms, and tools for visualizing benchmark results through regret plots and win-rate matrices.
