This isn’t meant to be an attack, but almost everything you say here is false.
> PIDs are great but notoriously hard to tune. They require deep insight into the underlying physical phenomena to get right. They are also rather rigid and cannot adjust well to a changing environment (temperature and humidity can fluctuate dramatically between summer and winter in some climates).
This is not true. PID controllers are often the least dependent on the physical characteristics. They can be tuned with heuristic methods like Ziegler-nichols, often with no knowledge of the actual system.
> Of course, no one tunes them by hand anymore for these reasons, relying instead on optimization techniques like particle swarm to find the best set of coefficients for a given steady state condition.
This is also not true. In the Amazon consumer robotics group we still tuned pid by hand. I’ve _never_ heard of tuning pid with particle swarm, that seems very silly, difficult, and overkill. If you’re going to use an optimization technique, you might as well move to a better controller structure like LQR. I have seen particle swarm used as an estimator, as an alternative to a kalman filter, but never seen it used for tuning.
> Eventually, I suspect we will replace most PIDs with a small neural network for almost all industrial applications (a handful of nodes is sufficient). The neural network is also easier to adapt to changing conditions.
This sounds unlikely to me. Classic control techniques give guarantees that a neural net just can’t. For example, things are provably stable under some assumptions. With a neural net you get no such guarantee. Also, it would be harder to debug and understand, and it would take more memory and compute. I can’t imagine a world where we replace pid with neural nets, they’re fit for very different purposes.
Source: have a masters in controls, worked in robotics in controls team, still do consulting in this area when I have time, and I love it all.
I saw a crazy stat somewhere like some large fraction of the industrial PID controllers that are actually deployed in plants are completely untuned--that is, just running default settings. I think it was a book about tuning PID controllers, but I don't have the exact citation handy..
I’d totally believe that. You can buy devices that even tune themselves automatically, using model-free methods like Ziegler-nichols. I’m not familiar with the domains where you would use that sort of thing, but I guess it would be applications that aren’t safety-critical, maybe air conditioning or something like that. I bet many people put the thing in place, and then it works well enough that they never even click the “tune” button
I don't have a good intuition about the what makes a problem sensitive to tuning, is it possible there are a large number of control problems where proper tuning doesn't matter much? Or the deleterious effects of improper tuning (like oscillations or overshoot) are masked somehow?
EDIT: I guess intuitively, big (lots of inertia) damped systems are probably pretty safe--you can do all kinds of crazy things with the control input and it won't really have much effect. The only way you could go wrong is drift.. Anything that is inherently stable seems like it should be "easy"--like a high-wing monoplane with lots of dihedral angle, you release all control inputs and it defaults to straight and level flight.
Yes, you’re absolutely right, systems that are default-stable are much easier to tune, and large inertia does make the system more stable. For those sorts of systems, any minimal feedback is often enough to get the performance you want. If you have high requirements you might need something fancier, but frankly most systems don’t have high performance requirements. Having very fast feedback loops helps too, so with digital controllers getting better and better you can be a bit sloppier with control design and it will still work.
There are indeed a large number of control problems where proper tuning doesn't matter much. I think we’ve built many of our tools to be “easy” to work with, and one aspect of this is that they’re intentionally made in a way that’s easy to control. Another factor here is that the “difficult” problems need some serious thought, which require research, measurement, and advanced degrees, making them more expensive. Many of these are just not worth the cost (yet). And even if you _do_ design a well-performing complex controller, you need to hire controls engineers to maintain and update it as designs change. I _love_ using LQR, optimal control, robust control, etc, but can almost never justify it. As a result, probably 90%+ of control applications by count just use PID. The remaining 10% are of course where most of the research happens, they’re much more fun.