Batch-sequential active learning for expensive computer experiments, together with learning-curve extrapolation for forecasting the remaining evaluation budget. Provides uncertainty-based and pool-based query strategies (greedy sampling, query-by-committee, inverse-distance-based active learning, and others) built on the 'bbotk' and 'mlr3mbo' optimization framework, and callbacks that record held-out surrogate performance after every batch. The resulting learning curves are represented as ordinary 'mlr3' tasks with dedicated learners (parametric curve fits, monotone smoothers, residual-bootstrap, conformal, and forward-simulation methods), point and distributional measures, and batch-wise expanding-window resampling, so extrapolation methods can be benchmarked like regression learners. Trained extrapolators can be turned into forecasts of the number of additional batches needed to reach a target performance.
Author
Maintainer: Martin Binder mlr.developer@mb706.com