Prof. Anima Anandkumar

Here is the line of work on partially observable linear dynamical systems (LQG settings): from T^2/3 regret https://arxiv.org/pdf/2002.00082.pdf …, to T^1/2 with absolutely new parameter estimation method https://arxiv.org/pdf/2003.05999.pdf …, to our finally log(T) regret bound https://arxiv.org/pdf/2003.11227.pdf ….https://twitter.com/AnimaAnandkumar/status/1263346130399580161 …

Btw, the log(T) is cool, but is not this one of those instances where the variance kills the log anyways? In other words, mean is small, but variance is just big.. Same comment applies to finite armed bandits. Just thinking aloud.

Ya, in bandit the var can kill, even can be linear in T. For LDs, I havent fully understood the behavior of var yet, long been thinking about it. For the logT, we dont encounter this issue since reg is defined as "sum c_t(\pi_t)-sum c_t(\pi^*)", under same realization of noise.