What's new in Ax 0.2 release? Master issue for announcement and discussion

[lena-kashtelyan]

The 0.2 release was quite a major one with a few brand new features. Please refer to this issue for information and of course please ask any questions!

1. Modular BotAx + tutorial: new flexible setup for combining BoTorch subcomponents like a Model, AcquisitionFunction, AcquisitionObjective, etc. into a Model in Ax (which can then be used with any Ax API). There was some discussion about the alpha release of this functionality in Modular BoTorchModel: combining BoTorch components like Model, AcquisitionFunction, Objective etc. in Ax #363, and we've revised it since to make it more convenient.

2. Scheduler + tutorial: new standalone configurable manager abstraction to run asynchronous closed-loop experiments with Ax (with or without human-in-the-loop). This functionality will be of interest to anyone looking to run full Ax experiments with automated deployment of concurrent trials on remote servers or local processes (like in Jupyter notebooks) without the need for human oversight. There is a short example on the API documentation page.

3. GenerationStrategy tutorial: walkthrough for optimization algorithms specification in in Ax; this has been a subject of many Github issues : ) Generation strategy allows our users to specify which optimization algorithm they want Ax Service API (AxClient) to use, to run experiments with Scheduler, and much more.

4. Bayesian optimization with categorical kernels exposed in Ax and auto-selected in Service and Loop APIs (via choose_generation_strategy; this modeling setup will appear as BO_MIXED in default generation strategies). This brand-new methodology eliminates the need to opt for Sobol quasi-random search when a given search space contains predominantly choice (a.k.a. categorical or discrete) parameters (unless there are too many combinations of categorical values in the search space, in which case currently Ax will still default to Sobol –– this is a temporary limitation of our implementation and will go away in the future). More documentation on this methodology will be added in subsequent minor releases.

5. ScalarizedOutcomeConstraint support (feature request: ScalarizedOutcomeConstraint support (e.g. sum outcome constraint) #472),

6. Extended exp_to_df utility to conveniently display an Ax Experiment as a dataframe of trials and arms (there is a usage example in this section of the Scheduler tutorial).

[lena-kashtelyan]

Closing this as it's been around for a while by now