Fuzzy PID has no zAligned x Medium rule

[dberlin]

Describe the bug

While increasing test coverage on my branch, i discovered an annoying bug in fuzzy PID.

create_fuzzy_pickle() has no rule pairing delta["zAligned"] (at setpoint) with current["MediumLow"]/["MediumHigh"] (the normal ~210-350F cook band). With no rule firing, skfuzzy's centroid defuzzification falls back to the near-empty aggregate, producing a very high hold cycleratio (~0.33-0.57) instead of a low near-hold value (which should be 0.077-0.3).

This causes fuzzy to perform ... quite badly ... at these temperatures in simulation and elsewhere

Measured effect on cycle ratio (setpoint == current, delta=0, rate_of_change=0):
225F: 0.358 -> 0.077 250F: 0.330 -> 0.089
300F: 0.570 -> 0.278 320F: 0.571 -> 0.278
No other input region changes; the new rules only fire where nothing
fired before.

Fix is easy:

   rules.append(ctrl.Rule(delta["zAligned"] & (current["High"] | current["Low"]), cycleratio["Tiny"]))
+    rules.append(ctrl.Rule(delta["zAligned"] & current["MediumLow"], cycleratio["Tiny"]))
+    rules.append(ctrl.Rule(delta["zAligned"] & current["MediumHigh"], cycleratio["Short"]))

[nebhead]

Thanks for raising this issue and for providing a suggestion!

Sounds like you're using the Fuzzy controller. With the fix, how's it performing for you? I frankly haven't touched this in a very long time as the PID's we have are pretty good. Just curious!

[dberlin]

Nah, this was noticed by improving coverage testing of branches/etc. There are other issues with fuzzy that are similar, and i don't use it, just noticed it and figured i'd report it.

Overall, I built an MPC replacement for the PIDs that performs much better than even the PIDs (<8F overshoot even under extreme conditions, and holds <1C RMS).
But it is more complex (it explicitly models auger->firepot deadtime, ambient loss, etc)

If you want an opinion, fuzzy should just be removed.

The overall hold RMS table looks something like this (I can include the overshoot and overshoot on temp change column if you want it, this is where MPC beats the tar out of PIDs ).

Before this particular fix, fuzzy was in the high 20's.

Hold RMS, C / mean temp

┌────────────────┬─────────────────┬─────────────────┬─────────────────┐                                                                                                           
│ controller     │ 220F            │ 375F            │ 425F            │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ MPC nonlin+MHE │ RMS 0.8 / 221F  │ RMS 0.9 / 376F  │ RMS 0.8 / 425F  │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ PID            │ RMS 2.1 / 220F  │ RMS 7.1 / 375F  │ RMS 8.5 / 425F  │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ PID clamp      │ RMS 1.7 / 221F  │ RMS 6.3 / 375F  │ RMS 7.5 / 425F  │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ PID clamp%PB   │ RMS 1.7 / 221F  │ RMS 0.9 / 376F  │ RMS 1.0 / 425F  │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ PID AC         │ RMS 1.9 / 220F  │ RMS 10.9 / 383F │ RMS 10.1 / 433F │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ PID-SP         │ RMS 9.6 / 236F  │ RMS 13.6 / 392F │ RMS 8.7 / 425F  │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ PID parallel   │ RMS 1.8 / 221F  │ RMS 6.3 / 375F  │ RMS 7.5 / 424F  │                                                                                                           
├────────────────┼─────────────────┼─────────────────┼─────────────────┤                                                                                                           
│ Fuzzy          │ RMS 14.4 / 238F │ RMS 13.4 / 386F │ RMS 11.3 / 427F │                                                                                                           
└────────────────┴─────────────────┴─────────────────┴─────────────────┘     

With a ton of fixes, you can get it to:

 ┌───────────────────┬──────────┬───────────────────┐                                                                                                                               
 │ metric            │ baseline │ wide_conservative │                                                                                                                               
 ├───────────────────┼──────────┼───────────────────┤                                                                                                                               
 │ 225→375 overshoot │ 29.8C    │ 11.0C             │                                                                                                                               
 ├───────────────────┼──────────┼───────────────────┤                                                                                                                               
 │ settle            │ no       │ 65m               │                                                                                                                               
 ├───────────────────┼──────────┼───────────────────┤                                                                                                                               
 │ 375F hold RMS     │ 13.4C    │ 2.5C              │                                                                                                                               
 └───────────────────┴──────────┴───────────────────┘     

But 65 minutes to settle is ... pretty crazy.

I tossed an agent at it for a few hours doing simulation and testing, and the agent's conclusion is:

No anticipation / deadtime handling
Rule-only fuzzy still dumps too much fuel during large setpoint steps. Conservative max limiting helps step response, but hurts high-temp recovery/hold.

My recommendation: implement the wide/soft near-setpoint surface first, then add a separate ramp-up limiter or step mode for large target changes. Rule changes alone probably won’t match MPC or PID.

Honestly, the complexity of getting fuzzy into a state where it competes with the PIDs or MPC is probably never going to be worth it.
There are better approaches at this point, and if it was me i'd remove it rather than try to support it.

[nebhead]

This is extremely impressive testing data! Kudos to you for doing all that work. Really cool. I'd love to see your controller make it into the official release some day.

I implemented Fuzzy a while back to just see how it would fare against PID, etc. I was mainly just trying to play with new options for fun and excitement. I only really shared it because, well, it was a thing that people might find interesting. I think I'll let it sit there, but maybe warn folks that it's quite terrible. Maybe someone would like to play with it some more some day.

Thanks again!