Kalman Filters and Learned Estimators
How classical model-based filtering relates to and combines with data-driven state estimation.
The Kalman filter
The Kalman filter is the optimal state estimator for a linear system with Gaussian noise. It maintains a state estimate and its covariance, predicting forward with a dynamics model and correcting with each measurement. Extensions (extended and unscented Kalman filters) handle mild nonlinearity.
Strengths and limits
- Principled, interpretable, and gives calibrated uncertainty
- Needs a known dynamics model and noise statistics
- Degrades when the model is wrong or the system is strongly nonlinear
Where ML enters
Learned estimators relax the need for an explicit linear model, capturing complex plasma dynamics and sensor behavior from data. But they can lose the clean uncertainty accounting and interpretability the Kalman filter provides. Neither is strictly better; they trade different things.
Hybrids
The productive middle ground uses each for what it does best. A network can supply the nonlinear dynamics or the measurement map inside a Kalman framework, keeping the filter's uncertainty machinery. Or a learned residual can correct a model-based filter's errors. These hybrids keep interpretability while adding the flexibility of learning.
Choosing
When a good physics model exists, a model-based filter is often the better, more transparent choice. When the dynamics are poorly modeled or the measurement map is complex and nonlinear, learning earns its place. For plasma state estimation, the answer is usually a hybrid, matched to how well the relevant physics is known.