Michael Dennis
- 84 Followers
- 115 Following
- 23 Posts
PhD student at the Center for Human-Compatible AI at UC Berkeley working on Game Theory and Reinforcement Learning
@akbir my guess is that people noticed we don’t have to include it for the marh to work out, and it allows us to transfer to a test time settings that doesn’t have the reward signal
Dylan Hadfield-MenellIf anyone here on #SigmoidSocial is interested in understanding the effects our #AI or #RL systems have on society, I would encourage you to check out the communities Tom is organizing (two of which are linked below)
The design decisions and research directions of today, are fixtures and industries of tomorrow. It's always great to find spaces where these considerations are discussed openly and seriously
[https://sigmoid.social/@tkgilbert/109547206519360129 | CHI workshop]
[https://sigmoid.social/@tkgilbert/109538077146112536 | PERLS reading group]
tkgilbert (@[email protected])
Do you have policy ideas for societal-scale AI? AND want to design for new capabilities? Consider submitting to our #CHI2023 workshop: "Designing Platform Technology and Policy Simultaneously". More info here: http://designpolicy.one/ We are soliciting short position papers (1-2 pages) and short research papers (4-6 pages). Deadline is February 23, 2023--plenty of time to put something together!
tkgilbertDo you have policy ideas for societal-scale AI? AND want to design for new capabilities? Consider submitting to our #CHI2023 workshop: "Designing Platform Technology and Policy Simultaneously".
More info here: http://designpolicy.one/
We are soliciting short position papers (1-2 pages) and short research papers (4-6 pages). Deadline is February 23, 2023--plenty of time to put something together!
@natolambert @bamos the elephant app brining people together 🐘🤗
@bamos low dimensional or high dimensional manifolds? There is a lot of computational geometry stuff I used to think about on manifolds, but all of it that I know of is exponential in the dimension and so stops working after 3 or 4.
Unfortunately haven’t been combining the two much yet! Just a Geometry -> RL convert. I think there are a good number of intuitions that transfer, or at least provide nice visualizations, but still looking for the “killer app” 🙂
Brandon AmosI am a scientist at Meta AI in NYC and study machine learning and optimization, recently involving reinforcement learning, control, optimal transport, and geometry. On social media, I enjoy finding and boosting interesting content from the original authors on these topics
I made this small animation with my recent project on optimal transport that connects continuous structures in the world. The source code to reproduce this and other examples is online at https://github.com/facebookresearch/w2ot
@bamos I also do Geometry/RL, thought I was alone! 😮
What kind of Geometry? :)
I mostly did discrete, 2-D geometry, usually around the Delaunay triangulation
Reminds me of:
DeepRL that matters
DeepRL at the Edge of the Statistical Precipice
IPPO and MAPPO
Measuring and Characterizing Generalization in Deep RL
The teammate transfer results in PSRO
(Probably many others, LMK if I’ve forgotten something!)
Seems like a reoccurring theme that should be taken more seriously if we care about real sustained progress in MARL
[https://twitter.com/michaeld1729/status/1604855967879172096?s=46&t=wxGlb03ZPSMgGiSpSUli_g | XP]
Very interesting result, showing that one can do very well on the very popular SMAC benchmark without looking at the observations!
This should be a reminder that progress in RL and MARL is hard to measure well, and even when progress is being made by a well defined metric, it is often not for the reasons we think!