Optimizer

The Optimizer class orchestrates the active learning loop by coordinating the acquisition function and search strategy. It selects the most promising candidates for evaluation by first obtaining a candidate pool from the search strategy, then scoring candidates using the acquisition function, and finally selecting the top-k candidates for oracle evaluation.

class alf_core.optimizer.optimizer.Optimizer(acquisition_fn, search_fn)

Bases: object

Implements the ask/tell interface for active learning.

Coordinates search and acquisition functions to propose candidates and update the surrogate model iteratively.

ask(state)

Propose the next batch of candidates to evaluate.

Uses the search function to generate a candidate pool, then the acquisition function to score and select the top candidates. In the first round, uses the training dataset if available.

Parameters:

state (State) – Current task state.

Returns:

• The proposed candidates to evaluate

• Updated state with ask_time metric

Return type:

tuple[list[Candidate], State]

get_metrics(state)

Collect metrics from acquired candidates, surrogate, and search functions.

Parameters:

state (State) – Current task state.

Returns:

• Metrics on acquired candidates (mean, max, min)

• Surrogate training metrics (if available)

• Search function metrics

Return type:

dict[str, float]

tell(state)

Update the surrogate model with newly acquired data.

Trains the surrogate model on the updated training and validation datasets, then computes and updates metrics.

Parameters:

state (State) – Current task state with updated dataset.

Return type:

State

Returns:

Updated state with tell_time, training_history, and optimizer metrics.