Hacker News Reader

Comment by Seden | Hacker News Reader

This is just a PID controller, they have existed and been used in industry much longer than computers and thus require almost no computing power at all. Everyone taking higher level engineering courses learns how to design these because almost everything uses them. Honestly I thought that the paper was a joke on how everything is called neural network, but maybe it is serious...

https://en.wikipedia.org/wiki/PID_controller#History

View on HN

It looks more like a cascade of a P controller (theta setpoint to gamma setpoint) with a PD controller (gamma setpoint to torque). But yeah, basically just classical control.

Yes, except that there is also a threshold applied (sigma).

Yeah, that was my thinking: how is this doing anything other than a PID controller? Glancing at the paper it actually looked like a PD controller, but maybe I missed some feed back where it could accumulate an integral term as well.

I guess you can think of it as a "self-tuning" PID controller. It "learns" how to ride a bike.

If it were really "just" a PID, you would need some arbitrarily complex procedure involving a human for tuning all the parameters?

It isn't self-tuning, and it doesn't "learn."

From the paper: "The three constants c_i need to be set by the implementor"

Under "Future Directions": "One obvious thing to do with this system is to have it learn and tune its parameters with experience." The work presented in the paper does not do this.

"This network was designed in an ad-hoc, if traditional, way. First, a human tried to control the bicycle with the simulator. After many attempts, the human finally became a somewhat skilled operator of the bicycle ... the human at this point was able to describe the key parameters which were being attended to, and based on this, the two-neuron network was designed." By "designed," the author means "the three constants were selected."

This sounds like the "arbitrarily complex procedure" you are looking for.

So it's a PD controller.

There's a whole body of theory on Adaptive Control (what the control systems community has called what we would now called "learning based control").

This is one of the classic texts: https://www.amazon.com/Adaptive-Control-Karl-Johan-Astrom/dp...

In short, many techniques that look a lot like modern machine learning have been in use for decades in control theory for tasks exactly like you describe (e.g. automatic tuning of PIDs). If you want to dig into this a bit deeper you might look up the "MIT Rule", which is a gradient style approach for dynamically tuning (and re-tuning) adaptive control systems, and is fairly common for PID tuning.

I'm with you here. Did no-one else do robotics in CS?

Unfortunately it's not a part of many curriculum. A lot of CS tracks seem to shy away from hardware.

For contrast, the bioengineering track at my state college did include electronics (along with programming, high level math and of course biology and chemistry).

Lot's of schools don't put robotics in CS. They put it in EE or ME. Computer vision is often a CS class, but control theory and other subjects that involve interacting with the physical world are often placed in other departments most CS students avoid.

I did as an undergrad, but I had to petition to take the class because it was considered graduate-level.

Funnily, I learned about PID controllers first while playing a Minecraft mod... but yeah, this definitely looks like one.

Reply on news.ycombinator.com