A Distributional Perspective on Reinforcement Learning · HackerPapers discussions

A Distributional Perspective on Reinforcement Learning — HN submission

Tue, 28 Aug 2018 20:33:16 GMT

A Distributional Perspective on Reinforcement Learning

1 point; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

A Distributional Perspective on Reinforcement Learning — mentioned by clickok

Fri, 06 Apr 2018 22:28:53 GMT

Yes. In particular, it's possible to learn the variance of the return using TD-methods with the same computational complexity as learning the expected value (the value function). See [0] for how to do it via the squared TD-error, or [1] for how to estimate it via the second moment of the return, and my own notes (soon to be published and expanded for my thesis) here [2]. It turns out that identifying states with high variance is a great way of locating model error-- most of the real-world environments are fairly deterministic, so states with high variance tend to be "aliased" combinations of…

Paper and discussion history. A mention is not an endorsement.

A Distributional Perspective on Reinforcement Learning — mentioned by backpropaganda

Sat, 29 Jul 2017 02:47:25 GMT

You can say RL is not hard when you manage to teach all aspects of the SOTA RL algorithm (https://arxiv.org/abs/1707.06887) to your class such that they are able to answer any question about it (not just implement it). Good luck teaching metric spaces to code monkeys. You guys are doing good work teaching tensorflow and algorithms/models researchers are coming up with, but are slapping those same researchers by disrespecting what they're working on now. Some humility would be wise.

Paper and discussion history. A mention is not an endorsement.