MATLAB Answers

Matlab implementation of AlphaZero

38 ビュー (過去 30 日間)
Mohammad Gohardoust
Mohammad Gohardoust 2020 年 1 月 17 日 21:12
コメント済み: John D'Errico 2020 年 1 月 19 日 15:57
Hi,
Just out of curiosity, I'd like to know if anyone knows of a sort of ready to use AlphaZero algorithm implemented in Matlab?

  2 件のコメント

Greg Heath
Greg Heath 2020 年 1 月 18 日 14:23
What is the meaning of the term "AlphaZero" ???
Greg
Mohammad Gohardoust
Mohammad Gohardoust 2020 年 1 月 18 日 16:40
Hi Greg, please see the answer by John D'Errico.

サインイン to comment.

回答 (1 件)

John D'Errico
John D'Errico 2020 年 1 月 18 日 15:52
編集済み: John D'Errico 2020 年 1 月 18 日 15:53
Now that I have some free time...
I'm pretty sure the answer is no.
Alphazero is a tool that was developed that could learn games like chess and go, merely by being given the rules and the end goal, to win. The tool would then play itself, trying different things, then learning how to become truly good at said game. It has achieved that goal to such an extent that in chess and go, (plus some others I cannot think of) it is now better than the human world champion by a considerable degree.
In fact, the human chess champion has typically had a rating on the order of 2800, sometimes flirting with 2900. The current reining chess code has usually been Stockfish, with a rating on the order of perhaps 3500 or so. The point being that a rating difference of 400 points of more means the higher individual will virtually never lose a game to the lower rated individual. So Stockfish should blow away the human chess champion, as we have seen it can do.
However, AlphaZero was able to dominate Stockfish, meaning that AlphaZero is sufficiently higher in rating than even the best A-B code.
I would note that Alphazero has NOT successfully been applied to bridge, my personal favorite mind sport. The difference in my eyes being that go and chess are games of complete information, whereas bridge is not. A large part of bridge lies in the realms of communication & even psychology. Bridge is not just pure probability. That makes it more difficult to master, not that I think it will not happen in the near future.
AlphaZero was developed by DeepMind.
As such it is proprietary to that company. However, there are now open source implementations of such a concept, for chess, in the form of Leela Chess Zero (LC0), and others. In the latest online computer chess championshoips, LC0 was in the top 2 ocmpetitors, though I don't recall the winner, and LC0 was still improving.
Again, none of them are based on MATLAB, at least that is my understanding, nor are there currently any such implementations that I have heard of. That does not mean there are none, written by some individual I have not yet heard about.

  2 件のコメント

Mohammad Gohardoust
Mohammad Gohardoust 2020 年 1 月 18 日 16:24
Thanks John for your answer. Added to your points, AlphaZero has been recetly employed for global optimization of quantum dynamics: here is the link to the article in nature.
John D'Errico
John D'Errico 2020 年 1 月 19 日 15:57
Yes. I imagine the general idea will expand over time, as people find new applications for it. One step closer to Skynet perhaps...

サインイン to comment.

サインイン してこの質問に回答します。


Translated by