0
$\begingroup$

I was reading Stanford's CS 229 materials on Linear Quadratic Regulation ( Lecture note 13, youtube Lecture 18, around time 36 min). And it mentions that:

" the quadratic formulation of the reward is equivalent to saying that we want our state to be close to the origin.For example, if $U_t = I_n $(the identity matrix) and $W_t = I_d$, then $R_t = -\Vert(s_t)\Vert^2-\Vert(a_t)\Vert^2$, meaning that we want to take smooth actions (small norm of $a_t$) to go back to the origin (small norm of $s_t$)."

but why is that? Why the norm of the origin is smaller? $s_t$ for any time point are just a vector of n dimension which can take any value, right?

$\endgroup$
  • $\begingroup$ The norm of $s_t$ is smaller the closer it is to the origin. $\endgroup$ – Yuval Filmus Feb 18 '18 at 7:35
0
$\begingroup$

Minimizing $s^TUs=s^Ts=\|s\|^2$ minimizes the distance from the origin, because the norm of a vector is its Euclidean distance (https://en.wikipedia.org/wiki/Euclidean_distance) from the origin:

$$\|s\|= \sqrt{s_1^2+s_2^2+\dots+s_n^2} = \sqrt{(s_1-0)^2+(s_2-0)^2+\dots+(s_n-0)^2}$$

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.