Batch Selection Strategies

Strategies for selecting multiple points from scored candidates. These are used by BatchProposer to construct batches for parallel evaluation.

Details

When batch_size > 1, selecting the top-k candidates by acquisition score may result in redundant evaluations (e.g., many points clustered around the current best). Batch strategies address this by diversifying the selection.

All batch strategy functions return a strategy function with signature:

function(candidates, scores, batch_size, surrogate, archive, search_space)

The strategy function returns integer indices into the candidates table.

Available Strategies

• batch_strategy_greedy(): Select top-k by acquisition score (baseline)

• batch_strategy_local_penalization(): Penalize around selected points

• batch_strategy_diversity(): Mix acquisition score with distance to selected