[AI Seminar] Fwd: AI Lunch -- Zhaohan (Daniel) Guo -- November 8
vitercik at cs.cmu.edu
Mon Nov 7 08:16:31 EST 2016
This is a reminder that this talk is tomorrow, Tuesday, November 8th, at
noon in NSH 3305.
---------- Forwarded message ----------
From: Ellen Vitercik <vitercik at cs.cmu.edu>
Date: Fri, Nov 4, 2016 at 9:55 AM
Subject: AI Lunch -- Zhaohan (Daniel) Guo -- November 8
To: ai-seminar-announce at cs.cmu.edu, "Zhaohan (Daniel) Guo" <zguo at cs.cmu.edu>
We look forward to seeing you this Tuesday, November 8th, at noon in NSH
3305 for AI lunch. To learn more about the seminar and lunch, please visit
the AI Lunch webpage <http://www.cs.cmu.edu/~aiseminar/>.
On Tuesday, Zhaohan (Daniel) Guo <http://www.cs.cmu.edu/~zguo/> will give a
talk titled "A PAC RL Algorithm for Episodic POMDPs."
*Abstract:* Many interesting real world domains involve reinforcement
learning (RL) in partially observable environments. Efficient learning in
such domains is important, but existing sample complexity bounds for
partially observable RL are at least exponential in the episode length.
Polynomial sample complexity bounds are prevalent for fully observable
environments, so we looked for a way to do the same for POMDPs. Generally,
polynomial sample complexity bounds for POMDPs are impossible, since
observations may give no information about the underlying dynamics. However
we can build on recent advances in estimating latent variable models using
method of moments, specifically confidence bounds on method of moment
estimators for HMMs. These methods quantify problem specific properties,
allowing us to give, to our knowledge, the first partially observable RL
algorithm with a problem specific polynomial bound on the sample complexity.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the ai-seminar-announce