Four tanks. Two pumps.
Changing dynamics.
Inspect PI, multivariable MPC and two compact kernel policies on a recognized nonlinear process model. Fresh policy rollouts alone did not fix this kernel adaptation’s tracking failures.
Synthetic Johansson equations · full four-level observations assumed · 288 held-out episodes · 20 minutes per episode · saved results, no plant connection
Two lower-tank targets
One confirmation episode per condition (seed 63001). The aggregate table uses all 12 episodes.
Held-out outcomes
| Controller | Mean RMSE, cm ↓ | Stage cost ↓ | Samples outside envelope | Mean episode p95, ms | Total solver failures |
|---|
What the comparison tests
Minimum-phase PI pairs each lower tank with its direct pump. Nonminimum-phase PI crosses those pairings, following the relative gain array. MPC predicts every future move using the known nominal model and enforces pump-voltage and rate limits across the horizon.
Both kernel policies share the same initial data, 64-state dictionary, preprocessing and 12 update rounds. One uses fixed random-action data; the other gathers fresh softmax policy rollouts and retains earlier samples.
This is a disclosed FKDPP adaptation. It replaces the original kernel-subset selection with fixed Nyström features, changes the task and reward, and evaluates greedy policies. One learning-seed sequence does not establish architecture-wide performance.
What changes in the plant
The controller continues to use the nominal configuration when a hidden change occurs. The experiment measures that failure mode; it does not implement configuration detection or recovery.
Any operational application would need measured valve/configuration feedback, authorized local admission, independent calibration, local interlocks and readback. Generic source and simulation results do not establish operating-site readiness.
Inspect and reproduce
Source, both policies, all tuning candidates, episodes and receipts ↗ · Benchmark report ↗
Protocol and selected tuning
The protocol was frozen before tuning and training. Training, validation and confirmation seeds are disjoint. The table retains every selected controller’s failures.
Protocol SHA256