facebookresearch/nevergrad

Features

• Objective Function Minimization - Finds the minimum value of a target function using a variety of gradient-free algorithms.
• Gradient-Free Optimization - Implements a wide array of adaptive, gradient-free algorithms to find the minimum of objective functions.
• Asynchronous Optimization Frameworks - Provides an asynchronous engine that decouples parameter suggestions from result reporting to support non-blocking trials.
• Hyperparameter Optimizers - Provides a system for tuning model variables across diverse search spaces to achieve optimal performance.
• Parallel Evaluators - Runs function evaluations across multiple workers using an executor to speed up the overall optimization process.
• Hyperparameter Tuning - Provides a framework for searching optimal model parameters across continuous, discrete, and categorical spaces.
• Hyperparameter Optimization - Minimizes functions using a combination of continuous and discrete hyperparameters by transforming them into a unified space.
• Gradient-Free Optimization Libraries - Functions as a comprehensive toolkit for finding function minima using adaptive algorithms without derivatives.
• Search Space Definition - Sets the search area for optimization using continuous, discrete, or mixed variable types to constrain the search.
• Automated Parameter Searches - Evaluates multiple candidate parameter sets across parallel workers to accelerate the discovery of optimal values.
• Distributed Search Executions - Supports parallel function evaluations by distributing parameter search across multiple workers.
• Gradient-Free Optimization - Finds optimal input values for a target function using adaptive algorithms that do not require derivatives.
• Parallel Objective Evaluations - Evaluates multiple parameter sets across different workers simultaneously to find the optimal value faster.
• Parallel Function Execution - Executes the objective function across multiple threads or workers to discover optimal parameters more quickly.
• Numerical Parameter Optimization - Allows specifying the search space using scalars, log-distributed values, or categorical choices to guide the optimizer.
• Pareto Front Optimizations - Maintains a set of non-dominated solutions to optimize multiple competing objective functions simultaneously.
• Unified Parameter Spaces - Transforms continuous, discrete, and categorical variables into a standardized representation to handle mixed domains.
• Asynchronous Ask-and-Tell Interfaces - Implements an ask-and-tell interface that separates parameter suggestions from result reporting to support asynchronous workflows.
• Asynchronous Parameter Coordination - Requests suggested parameters and reports resulting values asynchronously to facilitate parallel execution.
• Search Space Definitions - Allows defining the search space for hyperparameters using continuous, discrete, or mixed variable types.
• Multi-Objective Optimization - Simultaneously optimizes multiple competing performance metrics to find balanced trade-offs on the Pareto front.
• Optimization Benchmarkers - Executes experiment plans across functions and optimizers to measure performance metrics like loss and elapsed time.
• Result Reproducibility Controls - Seeds the random state of parametrizations to produce deterministic and repeatable optimization runs.
• Optimization Algorithm Benchmarks - Provides tools to integrate custom test functions and optimizers to evaluate new algorithms against baselines.
• Ill-Conditioned Function Optimization - Solves optimization problems involving rotated or poorly scaled objective functions using comparison-based methods.
• External Script Execution - Wraps non-native code as functions by using placeholders to inject parameter values into external source files.
• Black-Box Optimization - Provides the ability to optimize external scripts and non-native code by treating them as opaque black-box functions.
• Constraint Handling via Penalty Vectors - Manages search space restrictions by treating constraint violations as additional objective dimensions to be minimized.
• Discrete Domain Optimization - Minimizes objective functions over discrete domains using discretization of continuous variables or softmax transformations.
• Noisy Function Optimization - Finds the minimum of objective functions containing noise using specialized noise management principles.
• Optimization Algorithm Chaining - Runs multiple optimization algorithms sequentially and passes learned information from one process to the next.
• Optimization Constraint Enforcement - Restricts the search space using violation vectors or boolean constraints to ensure candidates meet specific criteria.
• Optimizer Inoculation - Enables the injection of known candidate values and their losses into an optimizer to guide the search process.
• Optimizer Abstractions - Provides a common API that allows users to swap different gradient-free algorithms without changing their objective function logic.
• Logic Decoupling - Separates candidate suggestions from result reporting using an ask-and-tell interface to decouple optimization logic.
• Asynchronous Task Executors - Uses asynchronous task executors to distribute function evaluations across multiple workers or threads for parallel optimization.
• Optimization Result Visualization - Generates regret plots and win-rate matrices to compare the effectiveness of different optimization algorithms.
• Computation and Optimization - Platform for gradient-free optimization.

Star history