Compatibility with Ax

Ax is a package developed by Meta for orchestration of Bayesian Optimisation. By utilising Ax, a large amount of orchestration and book-keeping is automatically handled, minimising developer overhead and providing users with an established and documented interface. While there are a large selection of tools available for maximising a GP, our ULS (detailed in usecase_offshorewind) problem does not fit directly into the framework. We do the following to deal with this:

• Define the Experiment Metrics to be the parameters of the response distribution (e.g location and scale).

• When a ModelBridge is instantiated using an ax Experiment it automatically sets up GPs to track the Metrics.

• We customise the Acquisition function in the ModelBridge so that it appropriately handles our GPs (Typical acquisition functions try to find the maximum of the GP, but we are interested in a different Quantity)

The following shows a conceptual overview of the key ax components involved.

This page provides more information about the ax Experiment and its components: https://ax.dev/docs/experiment/.