Add an acquisition function¶
An acquisition function scores candidates by their expected value to the search, trading off
exploitation against exploration. Subclass AcquisitionFunction to add a new scoring rule
(alongside the built-in Greedy, UCB, Expected Improvement, Thompson Sampling, and CoreSet).
Base class: AcquisitionFunction. Key method: __call__.
It scores the pool produced by a Search function; the Optimizer bundles the two and
selects the top batch. See Core Concepts for how one round flows.
Steps¶
Subclass
AcquisitionFunctionundertools/alf_tools/optimizer/acquisition_functions/.Implement
__call__to score candidates given the surrogate’s predictions and uncertainty.Drop it into an
Optimizerand run aTask.Add tests and run the build / pre-commit checks.
Reference notebook¶
Acquisition functions: implement a custom
AcquisitionFunction.