MPC Convexity requires linear dynamics about decisionmaking

[zsunberg]

I think you should say something like "A common approximation to make U(a1:d) amenable to optimization is to assume deterministic dynamics. Moreover, if these deterministic dynamics are linear, and the reward function is convex, the problem is convex."

[mykelk]

Addressing your first comment now. As for sampling a random transition for each state-action input pair, you are right that you can't do that explicitly for an infinite set. However, you can seed the random generator. It makes things deterministic, but it doesn't necessarily result in a convex formulation. You can also take your samples deterministically using say the 10th percentile from a Gaussian distribution if you want to perform well even with pessimistic predictions.

[zsunberg]

Right, regarding the second part, I just think you should change the "for each state-action input pair", probably by just referencing 9.9.2 and 9.9.3

[mykelk]

Sounds good. I'll post a new version tonight.