Talk:AlphaGo Zero

How does the algorithm of AlphaGo Zero differ from the algorithm of the earlier versions of AlphaGo?

How has the processing been changed? This is important for understanding how AlphaGo Zero could have become better than AlphaGo Master in such a short time. In other words, we need more information on the AI aspect, not just the Go aspect.

I have heard that this one uses just one neural network rather than the two networks used by the previous versions. Is this one network a value network, a policy network or something else (what?)?

During reinforcement learning, does this version only play against a current clone of itself rather than playing against all earlier versions of itself?

How does it select which board positions to start with during training? JRSpriggs (talk) 18:45, 28 October 2017 (UTC)

See https://deepmind.com/blog/alphago-zero-learning-scratch/ for a review. The Nature article [1] has more details, if you can get ahold of a copy. And, of course, lots and lots of commentary all over the web. ~ J. Johnson (JJ) (talk) 21:25, 9 November 2017 (UTC)

Future of Go addition

@Ashinpan: It's really quite simple. If you have WP:SECONDARY sources that talk about AlphaGo Zero and the Future of Go summit, cite them! We'll see what they say. If you don't, then there's nothing for Wikipedia to say. And if Wikipedia was going to say something, it's not going to be something useless like "Then, why did Alphago Zero not get a chance to play there? The answer can be only a matter of guess until Deepmind reveals further information."

As a side comment, your earlier claim was totally wrong, about DeepMind keeping a secret other version of AG in its back pocket in case Master lost against Ke Jie. My recollection is that as a requirement for the matches that any reasonable human would insist on, DeepMind is not allowed to "change" their AI once the tournament starts, in the same way that you can't pay a friend to substitute for you in the middle of a tournament. They're allowed to fix event-shattering bugs, and if they want to they can write some logic to detect an attempted replay of an already-lost game and know to mix things up rather than do the exact same thing, but that's about it. SnowFire (talk) 15:32, 16 November 2017 (UTC)

@SnowFire: This is exactly my point: there is no WP:SECONDARY sources that talk about AlphaGo Zero and the Future of Go summit, obviously because no one has asked Deepmind about it. Then, a Wikipedia statement like "Then, why did Alphago Zero not get a chance to play there? The answer can be only a matter of guess until Deepmind reveals further information" can push someone to get some more information from Deepmind, leading in turn to more information and better edits here. — Preceding unsigned comment added by Ashinpan (talk • contribs) 04:27, 17 November 2017 (UTC)

@SnowFire: You are criticizing something I have already retracted, so I have no comment here.— Preceding unsigned comment added by Ashinpan (talk • contribs) 04:27, 17 November 2017 (UTC)

@Ashinpan: Please read the linked policy page. If there are no secondary sources nor tertiary sources, material should not be included in Wikipedia. Let me quote from the page:

Wikipedia articles must not contain original research. The phrase "original research" (OR) is used on Wikipedia to refer to material—such as facts, allegations, and ideas—for which no reliable, published sources exist. This includes any analysis or synthesis of published material that serves to reach or imply a conclusion not stated by the sources.

If you'd like, we can suggestively mention the date and let readers draw their own conclusions. That's about it, however. Additionally, I stand by my earlier complaint: if we are going to use Wikipedia for advocacy / investigation, this is an incredibly boring question to point out. The answer is totally obvious, which is part of why no secondary source covers it. SnowFire (talk) 05:39, 17 November 2017 (UTC)

@SnowFire: You do have a point concerning the policy on original research. But the policy mentions only ``facts, allegations, and ideas." Therefore, would it be OK if I ask questions instead, like ``Did Deepmind have Alphago Zero already in hand when the Future of Go summit took place? If so, why did the Zero version not get a chance to play?"? In my opinion, a question is a question, none of facts, etc. — Preceding unsigned comment added by Ashinpan (talk • contribs) 09:00, 17 November 2017 (UTC)

@Ashinpan: Yes, a question is a question, not a fact. So you should not place your question into the article unless you find it raised in a secondary source. You may point out the fact that the journal Nature had received DeepMind's paper on AlphaGo Zero before the Future of Go Summit. But you should not add your own speculation or question into the article. --Neo-Jay (talk) 11:43, 17 November 2017 (UTC)

FYI I've started a similar discussion at the page for the match Rolf H Nelson (talk) 19:47, 24 November 2017 (UTC)

It shouldn't be in here or in the other article. It's speculation. We don't know, how the paper looked on 7 April. There might have been later changes or additions to it. -Koppapa (talk) 08:24, 26 November 2017 (UTC)

@Ashinpan: Two items for your consideration. First: please sign your edits by adding the "four tildes" – "~~~~" — at the bottom of your edits.

Second: you have approached this issue from entirely the wrong direction. Out of your own mind (?) you have come up with this speculation that AlphaGo Zero was being held in reserve in case AlphaGo Master couldn't beat Ke Jie (which is quite ludicrous, as Snowfire has explained), and then you looked for "facts" to support that speculation. However, that is not the way things are done here. As an encyclopedia we only present the state of human knowledge, as best determined by other authorities. If some issue has reached a point of notability we might address what is known about the issue, but if there has been even a smidgin of speculation on this it certainly is not notable.

The date Nature accepted the AlphaGo Zero paper has no relevance to anything except your fanciful speculation, and to "suggestively mention" it (as Snowfire suggests) is to suggest an incredulous view that does not reach even the notability of an extreme fringe. It amounts to a violation of a principle we call "point of view" (see WP:NPOV). It comes about when you start with an idea and then try to support it, instead of starting with what is known and then reporting that. ~ J. Johnson (JJ) (talk) 21:37, 26 November 2017 (UTC)

AlphaZero deserves his own article

Alpha Zero is on different domains (+ Chess and Shogi) and has significant impact on each of them. It will probably also change the field of machine learning - if only because of the possible hype that will follow in the aftermath of the new paper. Thus, it deserves it's own article. If anything, than AlphaGo Zero could be introduced as a premature Alpha Zero. Atarust (talk) 11:58, 6 December 2017 (UTC)

@Atarust: AlphaZero redirects AlphaGo Zero now. It can be changed to an independent article if enough reliable secondary sources that are independent of the topic provide significant coverage on it (see Wikipedia:Notability). DeepMind's article on AlphaZero was just published on arXiv yesterday. I have not found a qualified secondary source yet. But I believe that we can find qualified secondary sources very soon.--Neo-Jay (talk) 12:49, 6 December 2017 (UTC)

Yeah this is extremely notable, just made the article on it, there's already secondary sources flowing in at this time. Ethanbas ^(talk) 22:19, 6 December 2017 (UTC)

Reads like propaganda

Unfortunately the whole main article reads like progaganda for DeepMind and Demis Hassabis. Much of the claims are unverified, and a much wider ranger of citations are required. — Preceding unsigned comment added by 61.15.31.188 (talk) 11:19, 10 December 2017 (UTC)

You'll have to elaborate on what your objection is. The core (but not rigorously independently verified at this time) claim by DeepMind, that a minor variant of the AlphaGo Zero was able to achieve superhuman performance in chess and shogi with only knowledge of the rules, is being taken at face value by the mainstream media; since organizations like the BBC considers these claims credible (for what it matters, I personally also consider the claims 100% credible), and are reporting the results as fact, we follow their lead per WP:RS and also report it as fact at this time, in the absence of mainstream sources claiming otherwise. Another core claim of the current article is that this held by certain sources to be an impressive feat, given the lack of human-preprogrammed strategies and the lack of the opportunity to learn from expert gameplay; this too seems a well-sourced, credible and (so far) noncontroversial opinion worth documenting. Rolf H Nelson (talk) 00:41, 11 December 2017 (UTC)

Incoherent paragraph

In December 2017, a generalized version of AlphaGo Zero, named AlphaZero, beat the 3-day version of AlphaGo Zero by winning 60 games to 40, and with 8 hours of training it outperformed AlphaGo Lee on an Elo scale, as well as a top chess program (Stockfish) and a top Shōgi program (Elmo).

If a single AlphaZero program is competing in three separate disciplines, this needs to be far better explained. — MaxEnt 07:46, 29 October 2018 (UTC)

make paragarfh on nature

make 59.93.228.126 (talk) 15:59, 31 March 2024 (UTC)