Really excited to share RL Unplugged. A collaboration between researchers @DeepMind and @GoogleAI to make RL more assessable to researchers, and hopefully more useful in the real world.
RL Unplugged🔌: Offline RL benchmark that comes with both data and implementations of existing offline RL agents. It makes it easy to enter RL research and allows reproducibility.More agents in Acme will come soon.
Github: https://t.co/pCl2ulHtjD
arxiv: https://t.co/nVmO3kOXdX
New work shows that active offline policy selection (A-OPS) can accelerate policy development in real-world applications like robotics. A-OPS helps to quickly identify the best policy even when evaluation time on the robot is very limited: https://t.co/slBMBiIc5U 1/
Four years ago, we open sourced pysc2. Deep Learning + RL then went from solving some "simple" tasks using @StarCraft, to becoming GrandMaster a couple years later with #AlphaStar. Good memories & fun / humble beginnings!
Last year we presented #AlphaFold v2 which predicts 3D structures of proteins down to atomic accuracy. Today we’re proud to share the methods in @Nature w/open source code. Excited to see the research this enables. More very soon!
https://t.co/6uiV51Xly5
https://t.co/CLo7EKubBT
@NandoDF Finding the best policy using offline data only may have a low accuracy, but evaluating all policies online could be too expensive. With a BayesOpt approach, we improve upon the two methods with smart active online interactions, we call it active offline policy selection (a-ops)
Policy selection is key problem when applying RL to challenging real world environments. Excited to share this work to make it more tractable. Led by @ks_konyushkova and @yutianc.
How to decide which AI policies to launch in domains with real data and very expensive interaction, eg interaction with humans, industrial plants, robotics, and AB tests. https://t.co/2S64v5dAIu
I am excited to share our ICLR paper on autoregressive dynamics models for RL. Explicitly capturing conditional dependencies between state dimensions improves forward dynamics and reward prediction, yielding SOTA off-policy evaluation on control tasks.
https://t.co/xELsmbhYBE
Today our machine learning and distributed computing teams, along with Google Brain, are releasing Launchpad, a system for building distributed machine learning algorithms. Visit https://t.co/Ua86xIIx7s for the code, examples, and some distributed RL agents!
I am delighted to invite everyone tomorrow for the first RL Theory Seminar talk of 2021 by Andrea Zanette. Andrea will explain to us why and how batch reinforcement learning can be much harder than online RL. For details check out https://t.co/sijc7YN0Nz
Blog post on our latest experiments on visual model based reinforcement learning
https://t.co/XlS1aME66z
One of the most stable and flexible libraries that I ever worked on
https://t.co/8wD4F3Ur0F
with @msaffar3@danijarh@harinidkannan@chelseabfinn@svlevine@doomie
@fully_us I placed an order on Jan 16. Got a consignment number from TNT on Jan 20. Got an invoice from Fully on Jan 21. I cannot track the shipment. I called TNT, they have no record of collecting it. I emailed you on Jan 25 but haven't gotten an update. Can you help?
@Zergylord Totally. Agents can use their actions to perform interventions that will shed light on the causal structure of the environment. Humans do.
But usually agents don't bc a) it's a HARD exploration problem, b) environments and agent embodiments are limited often making it impossible.
Exciting #NeurIPS2020 Offline RL Workshop panel. Consensus from @EmmaBrunskill, @svlevine, @NandoDF: offline rl frees us to think about bigger problems than we have before. @JohnCLangford goes further: "Offline reinforcement learning might be the key to reinforcement learning."
@kelkar_uma@EmmaBrunskill@svlevine@NandoDF@JohnCLangford Learning policies from logged datasets, as opposed to online RL that learns policies by interacting with an environment (which can be costly, or dangerous in some situations).
This is EPIC!! It will go in history as one of the best #neurips talks of all times. @isbellHFh and colleagues 💜💙💚💛🧡❤️ on: Can’t Escape Hyperparameters and Latent Variables: Machine Learning as a Software Engineering Enterprise https://t.co/I8Y6oS1uY9