Discussion about this post

User's avatar
Bosque's avatar

If you repeatedly solve short-horizon planning problems via value iteration... is it MPC or DP? To me DP is an algorithm for finding globally optimal plans/trajectories via brute force. MPC is a choice about which problems your system is going to solve in order to do something useful in the world. I think that most people implement MPC using local optimization. I like to think about optimal control in terms of Maximum Principle vs Dynamic programming, aka local optimization vs global optimization, aka Soviets vs RAND corp :)

Expand full comment
Justin Bayer's avatar

But how do you "approximate the terminal cost in some reasonable way" for MPC? I can see how heuristics can be found in fairly easy problems. As soon as the problem becomes complex or we approach large T or infinite horizon problems, I guess you want to have heard of approximate dynamic programming.

Or am I missing sth?

Expand full comment
5 more comments...

No posts