2/3 important and deep nets can help. As @earnmyturns
noted, at least in causality, the definitions and tasks are available and clear, just read a standard textbook (e.g., causality) and you’ll see it.
-
-
Replying to @eliasbareinboim @CsabaSzepesvari and
What about all the recent papers by
@ShalitUri & co? https://scholar.google.com/citations?user=aeGDj-IAAAAJ&hl=en&oi=ao …3 replies 0 retweets 10 likes -
Replying to @zacharylipton @eliasbareinboim and
All the work with (online) RL is also doing causal inference. You can interpret interactions with the environment via a do operator. Interestingly, however, I haven't seen much value out of merging CI with RL. One exception is doubly robust methods for offline policy learning
4 replies 0 retweets 32 likes -
Replying to @dustinvtran @zacharylipton and
Sorry for the self-citation, but a lot is happening - e.g., off-policy method thr. causal reason.: https://goo.gl/yaHmeh ; Where to intervene? https://goo.gl/WPTyrs ; Counterfactual bandits: https://goo.gl/EQbwq7 ; Learn causal model from do-dist(): https://goo.gl/iz3Ey3
1 reply 4 retweets 70 likes -
Replying to @eliasbareinboim @zacharylipton and
Those are great references. Bandits, recommender systems, and experimental design have definitely seen a lot of success with causal inference. Do you know of any with a state space or episode with >1 time step, which is typical in RL?
4 replies 0 retweets 14 likes -
Replying to @dustinvtran @eliasbareinboim and
RL off policy policy evaluation and optimization given batch data is one take on this. Our lab thinks a lot about this, most recent to appear in NeurIPS work on combining NN & assuming ignorability for the multi-step setting https://arxiv.org/abs/1805.09044 .
1 reply 1 retweet 6 likes -
Replying to @AIforHI @dustinvtran and
A question from a future partner. To most folks in causality research the words "assuming ignorability" mean stripping a problem from its causal content and solving a standard statistical problem instead (
#Bookofwhy page 283). Must you really assume that?1 reply 1 retweet 10 likes -
Replying to @yudapearl @AIforHI and
1/2 In our engineered systems, we are pretty confident about that we know how actions are generated -- see diagram below. Disregarding statistical efficiency, the backdoor criterion applies and makes the estimation problem trivial from a causal perspective.pic.twitter.com/xo0XSpbsB2
3 replies 2 retweets 21 likes -
Replying to @CsabaSzepesvari @yudapearl and
In many cases even if we know how actions are generated, we need to have some sort of exclusion restriction to make valid inferences. Here is a paper from myself and
@deaneckles showing how to use noise from the 1000s of A/B tests companies run to do this https://arxiv.org/abs/1701.011403 replies 2 retweets 13 likes -
Replying to @alex_peys @CsabaSzepesvari and
Thanks, Alex, Yes, generalizing findings to new environments, settings, or pops is essential in most scientific explorations, and it was what Judea & I called transportability theory https://ucla.in/2FNnjLx (or https://goo.gl/StQkYB ). I need to read to understand the relation.
2 replies 0 retweets 6 likes
Replying to @eliasbareinboim @alex_peys and
totally agreed
10:13 AM - 28 Nov 2018
0 replies
0 retweets
0 likes
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.