Ax is Facebook's adaptive experimentation platform designed to automate the process of iteratively exploring parameter spaces to identify optimal configurations efficiently. The platform implements machine-learning guided optimization, currently supporting Bayesian optimization and bandit optimization as exploration strategies. Bayesian optimization capabilities are powered by BoTorch, a PyTorch-based library for Bayesian optimization research that Facebook also maintains.
The platform addresses a wide range of real-world optimization challenges through its expressive API. It handles complex search spaces, multiple objectives, constraints on both parameters and outcomes, and noisy observations. Users can request multiple designs for parallel evaluation, both synchronously and asynchronously, with support for early-stopping evaluations. The API abstracts away optimization complexities while providing sensible defaults, making advanced techniques accessible to practitioners without deep optimization expertise.
Ax is built with production deployment in mind, offering automation and orchestration features alongside robust error handling for large-scale real-world use. The platform is highly configurable, allowing researchers to integrate novel optimization algorithms, models, and experimentation workflows. Installation requires Python 3.11 or newer and is available via pip for OSX, Linux, and Windows. Optional extras support Jupyter notebooks, MySQL storage, tutorials, and development dependencies. For fully Bayesian models like SAASBO and SAAS_MTGP, users can install the fully_bayesian extra to access JAX and NumPyro backends.
The platform is licensed under the MIT license and maintains comprehensive documentation at ax.dev. The project encourages community participation through its issues page for questions, feature requests, and bug reports. Contributors are directed to the CONTRIBUTING file for guidelines on helping develop the platform. The repository recommends installing from source with all optional dependencies for development work. Ax is cited in academic literature, with a 2025 AutoML conference paper documenting the platform's design and capabilities.
