Uncertainty-Aware ML Control
Control policies that quantify their own uncertainty can defer to safe fallbacks when they encounter conditions outside their training experience.
Why control needs uncertainty
A learned controller trained on a limited set of conditions can behave unpredictably when the plasma enters an unfamiliar state. Uncertainty-aware control equips the policy with an estimate of its own confidence, so it can recognize when it is extrapolating and hand off to a conservative, physics-based fallback.
Sources of uncertainty
- Aleatoric: inherent noise in measurements and plasma behavior
- Epistemic: model uncertainty from limited or unrepresentative training data
- Distribution shift: the current state lies outside the training region
Epistemic uncertainty and distribution shift are the dangerous ones for control, because they signal that the model has not learned the current situation. Ensembles, Bayesian networks, and distance-to-training-data measures estimate them.
Acting on uncertainty
When confidence is high the learned policy acts; when uncertainty crosses a threshold the system falls back to a simple robust controller or to safe limits. This safe fallback architecture bounds worst-case behavior, which is essential when a control error can damage the machine or trigger a disruption.
Calibration
Uncertainty is only useful if calibrated: an 80 percent confidence should be right about 80 percent of the time. Miscalibrated models either take reckless actions or defer so often they add no value. Calibration is checked on held-out conditions and monitored during operation.
For control of concepts like the Hyperion breeder, uncertainty-aware policies are developed in simulation with fallbacks that respect operational limits, since the plasma is modeled and the hardware is not yet built. The uncertainty reflects computational confidence, and safety-critical control ultimately relies on validated, bounded behavior rather than on the learned policy alone.