Zoom Logo

CBBS MS-18 - Shared screen with speaker view
Wasiur KhudaBukhsh
38:52
@Ankit: any error estimates?
Wasiur KhudaBukhsh
39:25
@Ankit: isn’t the action space infinite for your reinforcement learning? How do you handle that?
Ankit Gupta
42:24
Hi Wasiur…We don’t have any rigorous error estimates but we can plot the loss function trajectories for the validation data to see if the neural network model has been trained successfully.
Wasiur KhudaBukhsh
44:07
Thanks!
Ankit Gupta
44:28
The action space is a mapping emulated by a neural network. So we don’t need to explicitly handle the infiniteness of the states-space.
Hyukpyo Hong
01:03:14
Thank you for the interesting talk, Greg. You mentioned that S and I in the SIR model are trackable. Could you explain why the species R is not trackable? Intuitively, it appears that R can be 'tracked' (actually every single individual in a system) as there is no birth event in the system.
Grzegorz Rempala
01:27:52
I meant to say that SIR are all trackable
Grzegorz Rempala
01:29:38
But R are trackable trivially
Hyukpyo Hong
01:30:45
I see, thank you for the answer!
Grzegorz Rempala
01:30:45
B/c R are not changing state
Jaewook
01:33:09
@Jinsoo. Can you apply your approach to a rare event , e.g. extinction? In that case, the system operates with two time constants, one for convergence to quasi-stationary distribution and another for leakage to extinction.
Lea Popovic
01:46:27
@wasiur: very nice talk, I will get your paper off the arXiv and email you with questions (had to leave before end). Thanks!
Jae Kyoung Kim
01:48:22
Thanks Wasiur for the great talk. Could you explain FPT approximation formula one more time….if time is allowed…..?