New: four-tank PI / MPC / kernel policy comparison → · Local telemetry audit kit ↗
Two pumps. One pressure target.
Explore compact learned controllers under demand changes, actuator mismatch and telemetry faults. Every curve comes from a saved experiment, with conventional control as the reference.
Synthetic normalized process · 12 held-out episodes per condition · 12,800 training transitions · no plant connection
Pressure tracking
One fresh confirmation episode per condition. The table below aggregates all 12 episodes.
Measured trade-offs
| Controller | Mean RMSE ↓ | Energy proxy ↓ | Pressure excursions | Rejected proposals | Unknown readbacks |
|---|
Local authority to apply a setpoint
A proposal binds the model, operating context, observation quality, time, units and exact setpoints. The local controller checks admission and reports the outcome.
These examples illustrate the tested executable specification. This static viewer does not issue commands.
On-prem placement
Ramp, bounds, permissives, replay and readback
Model, proposal, context and approval
Setpoint proposals remain inside the authorized OT zone.
Generic design informed by architecture references. The one-way path carries northbound telemetry; it provides no command return route. This diagram is not a site drawing or an approved deployment.
Reproduce the evidence
Training and validation seeds are separated from confirmation. The original comparison ran 288 episodes. After wear exposed model mismatch, the updated comparison used fresh seeds: five controllers, 360 episodes and 50,400 supervisory decisions. Three saved model artifacts passed 16 decision reload probes each; the trained C99 predictor passed 128 numerical comparison probes.
Model, code, raw results and reproducibility receipts ↗
The public repository records each local run and independent included-compute cloud replay separately. Reload and C checks verify narrow numerical paths; they are not additional accuracy tests.
Primary papers and implementation scope
- Cui et al. — Factorial Kernel Dynamic Policy Programming, 2018
Factor-wise kernel preferences and weighted-softmax DPP targets. This adaptation uses a bounded dictionary and fixed offline rollouts. - Bokor et al. — Tube-Based MPC with Random Fourier Features, 2025 preprint
Implemented: compact learned RFF residual dynamics. This finite-shooting controller has no inherited tube robustness guarantee. The pressure-offset observer is a conventional empirical extension. - de Jong et al. — Kernelized Nonlinear DeePC, 2025 preprint
Next predictor comparison; not implemented in this release. - Bras et al. — RL adaptation of explicit MPC, 2026
Motivates unseen mismatch and negative tests; not replicated here.