Efficient Optimization with Ax, an Open Platform for Adaptive Experimentation (engineering.fb.com)

Meta has released Ax 1.0, an open-source (MIT) platform for adaptive experimentation that uses machine learning to automatically propose and evaluate promising configurations for costly, high-dimensional experiments. Backed by a new paper, Ax combines Bayesian optimization (via BoTorch) with Gaussian process surrogate models and expected-improvement acquisition functions to balance exploration and exploitation. It supports multi-objective and constrained optimization, sensitivity analyses, Pareto-front visualizations, and experiment orchestration tools so teams can both find optimal configurations and understand trade-offs across many metrics. Ax is significant because it operationalizes state-of-the-art black-box optimization at production scale, reducing the time and resources needed to tune models, infrastructure, compilers, and even physical designs. Meta reports broad internal use—hyperparameter/architecture search, GenAI data-mixture discovery, recommender-system constraints, AR/VR hardware design, and material mixes for low-carbon concrete—demonstrating real-world impact. The platform is extensible (surrogates, acquisition functions, integrations) and intended for researchers and practitioners; get started via pip install ax-platform and the Ax website. The release invites community contributions to extend methods and production integrations.