Some reflections on the learning-to-control renaissance.https://www.argmin.net/2020/06/29/tour-revisited/ …
-
Great post. There is also adaptive control of partially observable LQR, i.e. LQG, in
@beenwrekt’s work sqrt T regret is proven. Our recent work shows logarithmic regret via new closed-loop system id method and online learning https://bit.ly/2ZKTnGX@SahinLale@kazizzad@Caltech
Continuous exploration is achieved by noise in output observations, not present in LQR + continuous model refinement achieved by closed-loop system-id allows faster rates in LQG. In regret sense, adaptive control of LQG easier than LQR. https://arxiv.org/abs/2003.11227
