After selecting each point, penalizes nearby candidates before selecting
the next. Encourages spatial diversity in the batch.
Usage
batch_strategy_local_penalization(bandwidth = 0.1, penalization = Inf)
Arguments
- bandwidth
(numeric(1))
Penalization bandwidth. Points within this distance (normalized) of a
selected point have their scores penalized. Default is 0.1 (10% of domain
range).
- penalization
(numeric(1))
How much to worsen scores of nearby points. Added to acquisition score.
Default is Inf (effectively removes nearby candidates).
Value
A batch strategy function.
Details
Algorithm:
Select the best candidate
Penalize scores of candidates within bandwidth of the selected point
Select the next best (from penalized scores)
Repeat until batch is complete
Distance is computed using Gower distance (handles mixed types).
References
Ginsbourger, D., Le Riche, R., & Carraro, L. (2010). Kriging is well-suited
to parallelize optimization.
