Active Learning for Costly Experiments
Active learning chooses which expensive simulations or discharges to run next, extracting the most information per costly evaluation.
When each data point is expensive
High-fidelity simulations and plasma discharges are both scarce and costly. Active learning addresses this by letting the model choose which point to evaluate next, rather than sampling blindly, so the training budget is spent where it most improves the model.
Acquisition functions
The model maintains an estimate with uncertainty and selects the next evaluation by an acquisition rule. Common choices target the point of highest predictive uncertainty, the point expected to most reduce overall uncertainty, or, in optimization, the point most likely to improve the best result found so far.
- Uncertainty sampling: query where the model is least confident
- Expected model change: query where a label would most update the model
- Query-by-committee: query where an ensemble disagrees most
- For optimization: expected improvement over the current best
Loop
The workflow iterates: train on current data, score candidate points by the acquisition function, run the most valuable one, add its result, and repeat. Over many rounds this concentrates costly evaluations in the informative regions and reaches a target accuracy with far fewer runs than uniform sampling.
Practical care
Active learning depends on trustworthy uncertainty estimates; if the model is overconfident it will stop exploring and miss important regions. Batch selection, which picks several diverse points at once for parallel evaluation, must guard against choosing near-duplicate candidates.
For a program building expensive high-fidelity databases, such as neutronics or gyrokinetic scans for the Hyperion breeder concept, active learning maximizes what each simulation teaches. It is a computational efficiency method; the underlying evaluations remain simulations for a machine still in design.