Select

The Select model define adaptive experimental design algorithms that select the next computations based on predictions from the machine learning models.

Available Selectors

ExaMol includes selectors that have a variety of characteristics

Selector	Category	Multi-Objective	Batch Aware	Multi-Fidelity
RandomSelector	Baseline	✔	✘	✘
GreedySelector	Baseline	✘	✘	✘
ExpectedImprovement	Bayesian	✘	✘	✘
BOTorchSequentialSelector	Bayesian*	✘	✘	✘
EHVISelector	Bayesian	✔	✘	✘

* Interface to many classes of selection algorithms

Using a Selector

Selectors employ a batching strategy to work with very large search spaces.

Start the selection process by creating the Selector then updating it with the current database and objective functions (i.e., recipes).

selector = GreedySelector(to_select=2, maximize=True)
selector.update(database, [recipe])

The Selector can then receive new predictions as a list of “keys” that define which computation associated with a list of of predictions from an ensemble of machine learning models.

# Samples array is: <num objectives> x <num samples> x <num models>
selector.add_possibilities(keys=[1, 2, 3], samples=np.arange(3)[None, :, None])

Continue to add new possibilities until ready to select new computations, which are retrieved by the “dispense” function.

print(list(selector.dispense())) # [(3, 3.), (2, 2.)]

Call start_gathering() when done dispensing to clear any remaining results or call add_possibilities() to start adding new possibilities again.