Hacker News Reader

Comment by wildekek | Hacker News Reader

When your espresso machine has some thermal mass (eg E61 group head), then a simple bang-bang controller performs just as well as a PID. After tweaking PID controllers in drones, HVAC systems and my espresso machine, my experience is that PID's do more harm than good in systems with a lot of inertia and slow response.

View on HN

High inertia/slow response is challenging to control systems in general, especially to linear systems like pure PIDs (especially if your 'slow response' is well modeled as transport delay, rather than just... slow response).

Bang-bang controllers with deadzones/hysteresis are really the simplest form of non-linear controllers, and so can do okay at something like an HVAC system, especially when precise tracking of target isn't needed. But you can strap all sorts of non-linear things (gain scheduling, hysteresis, etc) onto the form of PID to try to get decent macro scale behavior, while still having reasonable fine tracking performance. Though obviously, it becomes much much more complicated to tune.

If you know characteristics of the system, like thermal mass, inertia etc., then coupling feed-forward control along with the PID can help. The idea is that if you know the thermal mass (and you do for a coffee maker's internal elements), then you have a reasonable estimate of how many watts are needed to change the temp of those parts. So, don't make the PID put those watts in, instead just add them separately based on your simple physics model. The PID runs in parallel and corrects all the short-comings of your model.

Thanks, I never really understood feed-forward well, but modeling the thermal mass makes a lot of sense for keeping the terms in check. (As I understood it in Betaflight, it overrides the PID controller and mixes setpoints directly into the output.)

I found it funny that if you did robotics this is the first thing they teach you, and in control systems they went onto other more "fancy" controllers, but never explained how to make PID work well for a range of systems.

Chances are, you’re probably tuning it incorrectly. Tuning is not an easy thing to do, but you should get much closer if you draw out the block diagram. At a minimum, even fuzzy logic would be better than bang on/off.

I'm familiar with algorithmic and manual tuning, and the more latency the system has, the more time consuming and harder it is to get right. Genuinely curious how a block diagram would help?

PID is the controller for a block diagram [1]. When you have the block diagram, you can predict the response graphs extremely reliably [2]. You should be extremely close to the desired coefficients and requires a little tuning. Especially for stuff like thermal and conduction that’s very well understood. You just need to draw the thermal circuit and know the spec of your hardware.

Most of the time people use PID, they use it as a black box which causes a lot of pain.

[1] https://www.tutorialspoint.com/control_systems/control_syste...

[2] https://ctms.engin.umich.edu/CTMS/index.php?example=MotorSpe...

Thanks! I wish I learnt more control theory in my education, this is helpfull.

Control system was probably one of the hardest class I took for undergraduate. Unfortunately, it really requires a college course (mine had a lab too so you can some “real life” experience) to understand the beauty of PID controllers. Highly recommend sitting in a college course if you like this stuff!

Here's the first of several open courseware on controls.

http://mocha-java.uccs.edu/ECE4510/index.html

I think you are confusing terms. You are saying latency, but I think you mean the time constant of the poles / zeros. When one speaks of latency in a control system, they general mean a real time delay, like a digital processing delay, communication delay, etc. When you speak of something having low-bandwidth, such as due to the thermal mass and heating element you are describing, then one would say it has a long time constant. I have never seen these systems any harder to implement with a PID provided you select the proper terms to keep it stable. If it is just a linear PID, it is going to need to be a slow loop in this case.

A PID allows you to reduce the thermal mass, what makes your machine turn on faster and wastes less energy on power cycling.

But yes, adding one to a non-controlled machine that is already working well will probably harm its performance.

They key phrase in your comment is "that is already working well".

After many years of fixing things that aren't broken, I've come to appreciate the phrase "If it ain't broken, don't fix it!".

> PID's do more harm than good

In what way?

PID controllers require more tuning, which becomes harder with more latency in the system. They are more susceptible to integral windup: https://en.wikipedia.org/wiki/Integral_windup They require an MCU, whereas a bang-bang controller can be done mechanically (or with analog components), keeping them simpler.

Since when does a PID controller need an MCU? You can build one out of op-amps, resistors, and capacitors. Whether you should is a different story.

(I’ve never been convinced that a PID controller really makes sense for a heater where there may not be much natural cooling going on, but that’s a different story.)

Yes, you're correct. I think you never see them in practice since MCU's are cheaper, but it is totally doable.

There might be other reasons for wanting a MCU but PID controller chips are cheaper and so are the diy PID circuits.

Drone PIDs are hard, and a lot of that to do with the quickly moving targets which is not the case in many (most?) scenarios PIDs are used.

> Drone PIDs are hard, and a lot of that to do with the quickly moving targets which is not the case in many (most?) scenarios PIDs are used.

You can write a flyable PID in a few lines, and a full PID in a few more. Set the error term to the range-mapped input - angular rate measurement, for each axis. (Assuming you have the control inputs, gyro/acc inputs, and motor output IO setup)

I'm talking industrial PID and other consumer goods. Drones are just one place PIDs are used, though probably the most frequent place for interaction for software engineers on HN.

As for them being hard for drones... PIDs are built into the software which is easy enough to write but the drone characteristic is specific to both the configuration of the drone and the flight preference of the pilot. It can be hard to find an optimal PID setting for both, which is where it seems that most people are finding PID settings to be difficult.

Indeed, most modern approaces to controlling drones use other more advanced techniques that (ab)use the computational power of modern MCUs, PIDs are used only near the actuators to keep the motor torque within an optimal range or things like that. For the flight trajectory fancy drones use things like model predictive control combined with convex optimization algorithms to avoid obstacles very reliably (see for ex. [1]).

[1]: https://youtu.be/mHDQcckqdg4

That video and the results are astounding.

And, circling back to your point about where to use the PID: I think that perhaps we need to split things more finely: The fancy part of that video was WRT overall flight trajectory planning. (See also your obstacle avoidance descriptor). You might still use PID for the angular rate control (or not), while feeding those target rates from a higher-level algorithm that commands attitudes, flight paths etc. (Or flight path -> attitude -> angular rates etc)

I went down a rabbit whole of constant angular jerk and time-to-correct targets for a while, but wasn't able to get it working.

So, there are multiple layers involved for something like in the video. A properly-tuned PID works well for manual angular-rate controls, and autopilot commands that reduce down to angular-rate controls. But PID tuning can be tricky; and/or can be done automatically; or replaced with something more sophisticated.

I would concur, especially with control devices that can turn on to heat.

Although, a lot of more basic controllers only handle binary states for input. I'd argue that a calibrated (bang-bang with thermistor/thermocouple) on an analog control would be much easier to tune and easier to implement. That way, the hysteresis could be dynamically tuned out of the system with a constant input to stabilize losses.

I did something similar to that with I experimented with a peltier heated 3d printing bed. I fried the first batch of peltiers since I wasn't aware they *HATED* binary states, and would die in a crack of the ceramic matrix. My next round was to implement "Using PWM to Generate an Analog Output" ( https://ww1.microchip.com/downloads/en/Appnotes/90003250A.pd... )

Reply on news.ycombinator.com