MuZero
MuZero is a Reinforcement Learning heuristic. It is the follow-up of AlphaZero. It has been conceived at DeepMind and University College London. The main change from AlphaZero is that the rules of the game to be learned are not explicitly provided but learned along the way by telling the algorithm if a move is allowed or forbidden and also telling it when a player won/lost.
It is the fourth iteration of this line of efficient heuristics:
1) AlphaGo - Learn to play GO from human plays.
2) AlphaGo Zero - Learn to play GO from zero without any human knowledge input.
3) AlphaZero - Learn to play any board game from zero without any human knowledge input.
4) MuZero - Learn without the rules of the game being explicitly provided see original publication on MuZero [1] (PDF[2]).
Python pseudo-code associated with the article is available here
A Python and Ray non-official implementation of MuZero is available here
See this article[3] or this post[4] for an overview of the original publication. Three Medium articles[5] describing the heuristic and the pseudo-code from the article have been written by David Foster.
References[edit]
- ↑ https://arxiv.org/abs/1911.08265
- ↑ "Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model" (PDF). Unknown parameter
|url-status=ignored (help) - ↑ "DeepMind's MuZero teaches itself how to win at Atari, chess, shogi, and Go". Unknown parameter
|url-status=ignored (help) - ↑ "DeepMind Unveils MuZero, a New Agent that Mastered Chess, Shogi, Atari and Go Without Knowing the Rules". Unknown parameter
|url-status=ignored (help) - ↑ Foster, David (2019-12-03). "How To Build Your Own MuZero AI Using Python (Part 1/3)". Medium. Retrieved 2020-01-03.
Category:Reinforcement learning Category:Deep Learning Category:Neural networks Category:Monte Carlo Tree Search Category:Games Category:Artificial intelligence
This article "MuZero" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:MuZero. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
|
This page exists already on Wikipedia. |
