Intractable Problems in Control Theory (1986) [pdf] | Hacker News Reader

Intractable Problems in Control Theory (1986) [pdf]

mit.edu

52 pointspizza11 comments

Very surprised to see this on HN front page.

Control theory is a fabulous subject and has nice relationships to computational theory because classical (and more modern [s/d/o/nl/MPC] control theory) are obsessed with optimality, reachability and provable optimality and stability. In practice most control systems in the world are not designed nor run under the assumptions necessary to have an optimal controller. This is a field with a tremendous gap between theory and practice.

I think given the hype and (hopefully) coming importance of automated cars/automation in general in our daily lives, control theory is going to become increasingly popular over time. Maybe kids 50 years from now will have it as a mandatory course in their CS program.

Control theory is already ubiquitous.

> Very surprised to see this on HN front page.

From the abstract and the first author's name, I guess this might be relevant to game theory? which tends to be a hot topic.

It's incredibly sad that control theorists spent so much time working out analytic solutions to dynamical systems, only to be so thoroughly beaten by Reinforcement Learning.

The recent successes of Alpha Go, Alpha Go Zero, as well work at OpenAI and Berkeley (especially on the incredibly physically-accurate simulator MuJuCo) show that the era of classical controls is dead. Indeed, it may be the case that seeking analytic solutions to problems like this is an artifact of a time when computation was expensive and human time was cheap.

In fact, recent work on robots learning to play with themselves from OpenAI [https://blog.openai.com/competitive-self-play/] suggests that in the near future, humans won't be needed for designing these algorithms at all!

well uh.. I mean.. I beg to differ. I guess my opinion is "if what we call machine learning is really an old dream, then what was called control theory (and still is called control theory) was actually the grandchild of the idea of machine learning itself -- and that present-day machine learning is maybe only becoming real 100 or so years after when it was first dreamable, and actually dreamed of"

That said, there are many problems that troubled the earliest philosophers that still trouble the latest philosophers.. ¯\_(ツ)_/¯ .. but computers do pose themselves as a fundamentally-new tool.

I really prefer the term Statistical Learning. It’s more accurate. I know, Computer Science guys hate it because it takes away a lot of the mystery and hand waving, but it more accurately describes what’s going on.

Is it an accurate description of neural net training?

Yeah, I'll disagree. I work in the field and classic control algorithms are incredibly pervasive and are in no risk of being supplanted. Linear quadratic estimators, Kalman filters, are still used everywhere. Extended Kalman filters are still the standard for navigation systems and GPS. Only a very slight change to linear quadratic estimators gives linear quadratic regulators and these are used just about everywhere. Power flow problems are solved using LQR controls, which is basically our entire power grid. Hell, even my Dad's grill uses simple control algorithms like a PID controller because it works well.

Look, I'm happy that machine learning is hot and it's an interesting research area with some good applications. That said, it's ridiculous to assert that all of the sudden that all of these mature fields are dead. Control algorithms will always have a place because these dynamical systems are based on classical mechanics, which are really accurate for an extremely broad section of the world that we care about. Machine learning algorithms build models using generic functions that don't have information about the underlying physics. Now, could there be hybrid approaches? Absolutely. However, till that point, having an algorithm that works directly on the kinematics of the model is a huge advantage and that's not going anywhere anytime soon.

> It's incredibly sad that control theorists spent so much time working out analytic solutions to dynamical systems, only to be so thoroughly beaten by Reinforcement Learning.

I couldn't disagree more. It's incredibly amazing that control theorists spent so much time working on analytic solutions to dynamical systems, so much so that they're ubiquitous and modern technologies depend on them.

> Berkeley (especially on the incredibly physically-accurate simulator MuJuCo)

Huh? MuJoCo was primarily developed by Emo Todorov and his lab at University of Washington, and now his company.

Reply on news.ycombinator.com