Using a given predicted quantile, or possibly multiple of quantile (e.g. median + (q.25 - median) * lambda) could be an interesting acquisition function to consider. For this, the "Surrogate" and "SurrogateLearner" machinery would probably need to be adapted, and maybe made more general (e.g. instead of requires_predict_type_se
just required_predict_type
with character(1)
type).