vizier

Why we included this project

Vizier is the open-source version of the optimization service Google runs internally, and it keeps the same study-and-suggestion design. You define a search space that can mix float, integer, discrete, and categorical parameters, ask for batches of suggestions, run your evaluations, and feed the results back so the service decides what to try next. That loop matters most when each evaluation is expensive, like training a neural network or running a long simulation, because you get more out of a limited budget. The Bayesian and evolutionary algorithms include a JAX-based optimizer, and the service can run as a distributed client-server system, letting a team share one tuning backend across many experiments and users. Researchers also get developer and benchmarking APIs for implementing and comparing their own algorithms without rebuilding the surrounding plumbing.