From the surge-margin example to an uncertainty-aware anti-surge controller
The question
Given its flow meter's real state-dependent uncertainty R(Q), does an active inference agent hold a centrifugal compressor's surge margin better than a fixed-margin anti-surge controller on the same plant? The example covers differential-pressure meters (orifice, venturi, nozzle) and Coriolis meters. "Better" is defined under the first result.
The continuation is open as a paper. Lead authorship is yours of course. I'm happy to co-author the meter derivation and the cpomdp side if that helps and my contribution is sufficient. If not I only ask you cite cpomdp.
What a first result looks like
One meter type, Coriolis. The example covers it, and it is the case where the epistemic term pulls toward surge.
Give it a measured or literature-sourced uncertainty model, zero-stability term included. The SEVA work below reports an uncertainty with every measurement and is one place to start. Put a hard surge constraint on the margin. Run the agent and a fixed-margin anti-surge controller on the same plant, with the same meter and the same disturbances. Sweep Λ for the agent and the margin setting for the baseline, and report both margin-versus-efficiency curves. Efficiency here is recycle avoided.
"Better" means the agent's curve dominates. At the same rate of surge-limit violations it holds a smaller mean margin. Or at the same mean margin it violates less.
A negative result counts. An agent that needs the hard limit to stay out of surge, and does no better than a fixed margin once it has one, is worth writing up.
Kind of work
- Application / industry example
- Validation against a benchmark or oracle (RxInfer, published data)
Skills that would help
- Flow metering or process instrumentation
- compressor or anti-surge control
- Bayesian filtering
- JAX
Where it starts
The example the question grows from is written up here as Anti-surge control, with its results and the script that computed them.
Respond
The discussion states the proposal in full, with what the software already provides, what is missing and the prior art. Reply there to take it on or to ask about it.
Discussion #192 by DanBoringName, last updated .