Papers of the day   All papers

Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model

Comments

Nov 21 2019 Ian Osband

This feels like a real breakthrough: https://arxiv.org/abs/1911.08265 Take the same basic algorithm as AlphaZero, but now *learning* its own simulator. Beautiful, elegant approach to model-based RL. ... AND ALSO STATE OF THE ART RESULTS! Well done to the team at @DeepMindAI #MuZero
5 replies, 802 likes


Nov 25 2019 Brandon Rohrer

I hope that MuZero, the latest work by @DeepMindAI, sets a trend for model learning in RL. It’s a powerful and largely unexplored middle ground between model-based and model-free RL. https://arxiv.org/pdf/1911.08265.pdf https://t.co/gj3n8qxgGs
0 replies, 140 likes


Nov 21 2019 Mark O. Riedl

MuZero (DeepMind) gets top performance on 57 different Atari games and matches the performance of AlphaZero in Go, chess, and shogi https://arxiv.org/abs/1911.08265 Does so without knowing the rules of the game, which it learns using model-based RL
1 replies, 123 likes


Nov 20 2019 Ankesh Anand

Exciting new paper from @DeepMindAI : Planning with **learned** models can scale to complex visual domains like Atari. TL;DR: MCTS + Q-learning with models that only predict rewards/Q-values leads to new SOTA on Atari games and matches AlphaZero on Go. https://arxiv.org/abs/1911.08265 https://t.co/d961KoFiiy
0 replies, 102 likes


Nov 21 2019 Noam Brown

"Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model" -- Really exciting new work from the AlphaZero group at DeepMind! https://arxiv.org/abs/1911.08265
1 replies, 83 likes


Nov 23 2019 Pascal Bornet

New @DeepMind's #MuZero wins at games WITHOUT prior knowledge of their rules https://arxiv.org/abs/1911.08265 #AI @alvinfoo @kashthefuturist @FrRonconi @Paula_Piccard @ronald_vanloon @jblefevre60 @evankirstel @mvollmer1 @HeinzvHoenen @samiranghosh @YuHelenYu @MHiesboeck @andy_lucerne https://t.co/wgAiejEyEa
1 replies, 67 likes


Nov 22 2019 Thomas G. Dietterich

I enjoyed reading this paper. Many nice design choices. I have some questions that mayydolks can answer. First, how hard is it to learn a really good transition model for chess? For go?
4 replies, 60 likes


Nov 21 2019 Adam Santoro

Another incredible step to make AlphaZero even more general, doing away with the simulator and instead embedding it in the learning loop. In before "but it still uses a highly structured tree search!"
6 replies, 44 likes


Nov 22 2019 Rob Miles

MuZero, the latest from Deepmind, can play Chess, Go, Shogi *and Atari*. I don't see a straight line from here to AGI, but it is remarkable how these systems keep getting more and more general. https://arxiv.org/abs/1911.08265
4 replies, 39 likes


Nov 20 2019 mooopan

Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model https://arxiv.org/abs/1911.08265 😲
1 replies, 38 likes


Nov 21 2019 Carlos E. Perez 🧢

DeepMind is so ahead of the curve. I was dreaming last night on a novel way to formulate an RL solution. Only find out this morning, that DeepMind implemented my dream and has a paper out! https://arxiv.org/abs/1911.08265 .
2 replies, 29 likes


Nov 23 2019 Darshan H Sheth ✨ @Iamdarshan

New @DeepMind's #MuZero wins at games WITHOUT prior knowledge of their rules https://arxiv.org/abs/1911.08265 #AI @pascal_bornet @darshan_h_sheth @alvinfoo @kashthefuturist @FrRonconi @Paula_Piccard @ronald_vanloon @evankirstel @mvollmer1 @HeinzVHoenen https://t.co/fLfMBYNDvP
0 replies, 28 likes


Nov 21 2019 Kaixhin

Alright, had to have a look at @DeepMindAI's SotA #MuZero: https://arxiv.org/abs/1911.08265 . AlphaZero, but with a learned model, leveraging good targets and predicting the policy, value and reward for rollouts to train representation, dynamics and prediction modules.
3 replies, 27 likes


Nov 21 2019 Simon Osindero

Agreed — this is great progress!
1 replies, 17 likes


Nov 21 2019 Santiago Ontañón

Man, one day not paying attention and I am already outdated! I was the only one not aware of DeepMind's MuZero in a discussion today! In case you also missed it: https://arxiv.org/pdf/1911.08265.pdf
2 replies, 13 likes


Nov 21 2019 Aran Komatsuzaki

Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model https://arxiv.org/abs/1911.08265 MuZero combines a tree-based search with a learned model and achieves superhuman performance, without knowledge of dynamics, at various games, including Go and Atari (sota).
0 replies, 8 likes


Nov 21 2019 Daisuke Okanohara

MuZero improves AlphaZero by using a learned simulator, which is trained for predicting reward, value, and policy only, Achieved new SoTA on Atari, Model-based RL surpassing model-free for the first time, and mastered Shogi and Go without the game rules. https://arxiv.org/abs/1911.08265
0 replies, 8 likes


Nov 21 2019 Santiago Ontañón

Very interesting work by @DeepMindAI on MuZero! They play Atari, Go, Chess and Shogi from pixels by learning a forward model from scratch for using MCTS (even to play Go, Chess, etc. they give the system a "picture of the board")
1 replies, 7 likes


Nov 21 2019 Pierre Richemond

MuZero - MCTS on Atari, chess-Go, SotA, !! no knowledge of game rules or environment dynamics required !! Key paper. https://arxiv.org/abs/1911.08265
0 replies, 5 likes


Nov 20 2019 Statistics Papers

Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model. http://arxiv.org/abs/1911.08265
0 replies, 5 likes


Nov 20 2019 Brundage Bot

Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model. Schrittwieser, Antonoglou, Hubert, Simonyan, Sifre, Schmitt, Guez, Lockhart, Hassabis, Graepel, Lillicrap, and Silver http://arxiv.org/abs/1911.08265
1 replies, 4 likes


Nov 20 2019 Eric Xu, PhD (徐宥) 🧢

Holy Cow this is a good step towards AGI. https://arxiv.org/abs/1911.08265
1 replies, 4 likes


Dec 01 2019 Dan Hughes

MuZero learned to play 60 games at superhuman levels without knowing their rules. MZ predicts the quantities most directly relevant to planning: the reward, the action-selection policy, and the value functions. https://arxiv.org/abs/1911.08265
1 replies, 3 likes


Nov 21 2019 Haruhiko Okumura

"When evaluated on Go, chess and shogi, without any knowledge of the game rules, MuZero matched the superhuman performance of the AlphaZero algorithm that was supplied with the game rules." https://arxiv.org/abs/1911.08265 Wow!
0 replies, 3 likes


Nov 20 2019 Federico Andres Lois

A new year, a new AlphaZero paper. Muzero now in model-free flavour without losing capability. #reinforcementlearning https://arxiv.org/abs/1911.08265
0 replies, 2 likes


Nov 22 2019 Barney Pell

Wow!!
0 replies, 2 likes


Nov 21 2019 Jonathan Raiman

Incredible work on Model-based RL that finally outperforms other approaches on both continuous/visual games (Atari) and board games (go, chess, shogu) from @DeepMindAI @Mononofu, Antonoglou, Hubert et al. So many problems can be cast this way, congrats! https://arxiv.org/abs/1911.08265
0 replies, 2 likes


Nov 30 2019 Yohan J. Rodríguez

#Tech #Automated | Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model https://arxiv.org/abs/1911.08265?utm_source=hackernewsletter&utm_medium=email&utm_term=data
0 replies, 1 likes


Content