Talk:Computer shogi

Latest comment: 6 years ago by Ish ishwar in topic Alphazero


Japanese Computer Shogi Article

edit

I see there is a link to the Japanese Computer Shogi article. Maybe somebody who reads Japanese can read that article and see if there is something interesting to add the English computer Shogi article. I guess this should be done on a regular basis, maybe twice a year. I see they list all the winners of annual CSA tournaments. I don’t think that is necessary. The CSA web page already does that and a link to there is sufficient. Mschribr (talk) 12:10, 25 June 2009 (UTC)Reply

Unprofessional wording

edit

Isn't "The match was not interesting because the 2 players were not in the same playing level" a matter of opinion that does not belong on a wiki page? In addition, the whole section about Watanabe vs. Bonanza seems to be written from one point of view rather than objectively. 68.191.84.156 (talk) 15:29, 14 December 2009 (UTC)Reply

You are right about the part that says the match not being interesting. I took out that part of the sentence. The rest of the section are either facts or backed up with references. Which sentences give only one point of view? Mschribr (talk) 17:06, 14 December 2009 (UTC)Reply

No amateur beaten in game longer than 2 hours

edit

This claim is made a couple of times in the article and I am unsure of what this is intended to mean. It does not make sense as written because I am certain that I can manage to lose spectacularly to a computer program. I suspect the idea is that no computer program is competitive against high level amateurs when the time limit is large. —Preceding unsigned comment added by 130.164.68.85 (talk) 22:28, 9 April 2010 (UTC)Reply

The article is trying to make 3 points.
  1. The computer won 1-hour games against amateurs. The next level of competition is a 2-hour game with an amateur.
  2. We do not know how a computer would perform in a 2-hour game against an amateur because these games were never played.
  3. The ban on playing pros is irrelevant because the computer has not beaten an amateur at 2 hours.
When the Wikipedia article says a 2-hour game, it means in a tournament game in a public arena and not a private game. Maybe Wikipedia should make it clear that these games are public tournament games. Mschribr (talk) 16:38, 13 April 2010 (UTC)Reply

Hidetchi

edit

Hi,

I don't feel that the Youtube reviews of "Bonanza vs. Watanabe" and "Akara vs. Shimizu" by Hidetchi cited in the article are appropriate to use as inline citations because of WP:UGC so I have removed them. This is not a reflection of Hidetchi's skill at shogi, but Youtube videos/commentary are self generated; Therefore, they could easily be considered to be original research and their verifiability questioned. Moreover, no mention of who Hidetchi is or why he or his commentaries are notable or essential to understanding any of the information they cite; So, this could also be seen as simply a way of promoting somebody's Youtube channel. These links may, however, be acceptable as external links per WP:YOUTUBE, and can easily be re-added as such. If you disagree with their deletion, please discuss. Thanks in advance. -Marchjuly (talk) 08:17, 17 July 2014 (UTC)Reply

Fullwidth/halfwidth characters

edit

Hi,

It seems that some of this article may have been edited using a Japanese keyboard. Of course, it's is very easy to input in English using such keyboards, but the default font for English is typically a Japanese one (for example, MS Mincho or MS Gothic on Windows computers) so some characters (particularly things like commas, quotation marks, apostrophes) are inputted as fullwidth (全角, zenkaku) characters instead of as the halfwidth (半角, hankaku) characters more common to British and American keyboards. No big deal really, just a minor formatting issue that would be hard to notice if you're not used to seeing it. Anyway, I've cleaned up the all of the fullwidth apostrophes I could find and then changed them to "typewriter apostrophes" per WP:PUNCT by using the template {{`}}. Hope this doesn't create any problems. - Marchjuly (talk) 01:17, 18 July 2014 (UTC)Reply

Miura vs. GPS

edit

I have removed the YouTube link for the Miura vs. GPS game per WP:YOUTUBE because it may be a possible copyright violation. The YouTube channel linked to is not an official channel of NHK World and, therefore, it is not unreasonable to assume that the person who uploaded the video is not the original copyright holder. Since the statement the link was supposed to be supporting is a direct quote, it should be supported by another more acceptable reliable source. I have added a {{citation needed}} template and will try and find a source myself, but the quote should be ultimately removed if such a source cannot be found or does not exist per WP:BLPSOURCES. - Marchjuly (talk) 05:00, 18 December 2014 (UTC)Reply

I added a source for the Miura quote with this edit . The text in question is located in the middle of the article and was given as part of the post-match Q and A. The Japanese text is as follows: 準備はしていたのですが、GPS将棋がこれほど強いとわかっていれば、もっと危機感を持って、より前からやっていればよかったと反省していますし、悔いが残るところです。 The translation given in the text was probably taken from the YouTube video and is accurate in my opinion, so I left it as is. - Marchjuly (talk) 06:09, 18 December 2014 (UTC)Reply

Reference cleanup

edit

I have removed the following sources and replaced them with {{citation needed}} templates because they link to discussion board forums or personal websites per WP:UGC or do not directly support what they are supposed to be supporting per WP:RSCONTEXT.

  1. Shogi Club 24 - This just links to the homepage of Shogi Club 24, which is regualry updated and, therefore, no longer supports any of the information it is supposed to be supporting. It would be much better to find third-party sources which are updated much less frequently and use them instead.
  2. "Yoshiharu Habu rates computer at the level of 2 dan shoreikai" - Original url is dead, but an archived version can be found here. Links to an archived online shogi discussion board thread in which Habu is mentioned, but source is incorrectly attributed to Habu himself. The article mentioned in the thread is what should be cited and attributed to Habu, not a email discussion about something Habu is supposed to have said.
  3. Watanabe comments on his game with Bonanza - original URL is dead, but an archived version can be found here. Link to an archived online shogi discussion board thread in which one of the forum's members posts his translation of something her read on Watanabe's blog. If anything should be linked here, it should be Watanabe's blog if that would be considered OK per WP:BLPSELFPUB
  4. Will the computer beat Yonenaga in Januarary 2012? - Another link to an online shogi discussion forum; Much better to link to the original source than some user-generated rehash.
  5. "Gsme Miura Hiroyuki vs GPS Shogi" - Links to another blog/personal website, no way to verify if the game score is copyrighted or not. Again, it would be much better to have a reliable third-party source cited.

In addition to removing the above, I also cleaned up the two sources for the "3rd Denou-sen". - Marchjuly (talk) 07:03, 18 December 2014 (UTC)Reply

just a note about the copyright bit: game records cannot be copyrighted in the US according to a recent chess lawsuit. Since the Wikipedia servers are mostly in the US, there should be no copyright issue regardless of whatever the copyright law in Japan is. (Incidentally, there was some new stories about this and the related issue of newspaper folks pressuring Youtubers to stop doing live personal commentary on in progress games.) – ishwar  (speak) 01:56, 8 December 2017 (UTC)Reply

Player profile pages

edit

I've removed the player profile pages on the Grimbergen website per WP:UGC and WP:LINKVIO. Not only were the links not really needed because links to the more frequently updated official JSA profile pages were also being provided, but the Grimbergen website contains photos of the players that might possibly be a copyright violation since some of the photos appear to have been taken directly from the JSA pages, e.g., Toshiaki Kubo, Masataka Goda, etc. - Marchjuly (talk) 07:52, 25 December 2014 (UTC)Reply

Denou-sen (2013)

edit

The section Computer shogi#Denou-sen (2013) is in need of much better sourcing and some rewriting. Here are some of the problems (in my opinion) which need to remedied:

  • The only source cited, Shogi Electronic Kings Battle channel is pretty much useless because the information on the Niconico page has been update and no longer says anything about "Denou-sen (2013)". The citations of the JSA shogi professional profile page's were also not very useful since they are general links related to the players themselves and not to their participation in this particular match. We don't need to verify that these professional are JSA professionals; We need to verify that they are the JSA professionals who participated in this match. The same goes for the software programs themselves. Only information in the table which can be supported by independent reliable sources should be left and everything else removed per WP:NOTEVERYTHING. All of the information about programmers, cpus, first tournament, etc. may be true, but they need to be verifiable.
  • The "rank" and "rating" categories are also irrelevant because they are not officially used by the JSA; They are the original research of others. If they are going to be used then it needs to be shown that use of these ratings/ranks has received significant coverage in independent reliable sources and are not simply a case of WP:UNDUE.
  • The use of a blog to cite the game scores is not really acceptable per WP:UGC. 2 Channel is nothing more than Internet forum with posts made by users and no semblance of editorial control or oversight and, therefore, would be considered unreliable. Again an independent reliable source (e.g., a newspaper or magazine, etc.) indicating who won (and possibly how many moves) is preferred.
  • The only table that is really needed is the "Games" table. The other two could easily be converted into prose per MOS:TABLE#Prose with the last table being retained as a summary of the section. It seems that the tables are just being used as space fillers and as a way to get around proper sourcing.

- Marchjuly (talk) 05:09, 25 May 2015 (UTC)Reply

time limits

edit

Recent interesting article: http://www.japanpolicyforum.jp/archives/culture/pt20160516000523.html

Morishita Taku suggests that one recent problem of underperformance are the time restrictions at the ends of games. If the time limits are increased to 15 minutes per move throughout the game, then the human error rate may decrease making humans at a more even playing level in the end game. He also suggested using a separate board for humans to work out moves during their time. (The implication is that humans may have some faulty memory problems especially under short time limits that computers don't have. And, ultimately, the games are meant to compare human reasoning vs a computer's objective function and not merely memory.) With these modifications, he was able to dominate the same computer program that beat him earlier. Don't know if it's worth putting in the article, but yall can consider it. – ishwar  (speak) 08:09, 19 May 2016 (UTC)Reply

If you are giving more time and a second game board for analysis to professional players then we need to remove the restrictions on the size of the computers. The current professional vs computer games greatly underestimate the power of the current shogi programs. Since 2014, shogi programs in man vs computer games have been restricted to single processor computers. If today's shogi programs were running on the same computers they were running in the World Computer Shogi Championship, the computers would win every game. In 2013, GPS beat Hiroyuki Miura, a player in the top 20 professional players. Instead of picking a stronger professional player to play GPS, they put the computer programs on weaker computers. So we do not see the full power of these programs. --Mschribr (talk) 11:39, 23 May 2016 (UTC)Reply
I don't see that as really all that related. But, they shoulda been had multiple processors. Let the programmers have all the resources available. And, they actually need to have many competitions. Just a handful of games isn't enough to establish statistical significance. Japanese folks are being weird here. (Obviously, it's partly related to the cultural importance of shogi.)
How do you know that the computers are so good? Are you saying that the practical search space is so small that with enough computing power they can exhaustively search the space? I don't know what they're doing, but some of the older papers I glanced at stated that they hard coded openings into the game play since the programs weren't able to perform well without it. That suggested to me that they're not doing something right. If they can't get the openings, how can they get the middle and end? The AlphaGo stuff makes a lot of sense in comparison: connectionist network optimized over human games which is further optimized with an iterative learning algorithm over simulated games. If the shogi folks are using something like that, then I think the humans should be scared (assuming that the things being optimized are the things needed to optimized, which is the hard part). But, if they're not doing that and using something that is built more by rules, then I can see the programmers failing. Unless the search space is small enough, which is my question. (And, if that's the answer, then it's pretty boring. I mean they can just ignore improving the objection function and brute force a win. That can be destroyed by moving to chu shogi or larger board with dropped pieces. At some point, they'll have to actually get the program right.)
But, back to the article, there isn't really any mention of the mechanics behind the shogi programs. It'd be good to have that in the article. (If it's not top secret...) The article is essentially: this won this, that won that, etc. – ishwar  (speak) 01:52, 24 May 2016 (UTC)Reply
I agree there should be more games. Only 2 games in 2016. I guess if you only play a few games then you only lose a few. Professional would look bad if they lost many games. The world computer shogi champion would beat any professional player. The shogi professionals vs computers games since 2014 are not a good indication of computer strength. In 2013, GPS beat Hiroyuki Miura, one of the top 20 professional players. That means GPS is stronger than Miura. But we do not know how much stronger, because GPS played only one game. We know that the number one program on the floodgate server increased a whopping 600 points from 2013 to 2016. So a program 600 points stronger than 2013 GPS, is more than enough to destroy any professional player. This helped Ponanza quickly defeat 8 dan Yamasaki in only 85 moves and Yamasaki used a lot of time, 7 hours 9 minutes. The search space is large. But the computers prune smartly and search deeply. Computers do not exhaustively search the space. If they did then the computer would play perfectly and never lose. The computer searches smartly, enough to easily beat the any professional players. They code opening moves to save time. They can hard code the opening because there are relatively fewer opening positions than middle game positions. Hard code the opening moves means you have more time to search deeper in the middle game and produce better moves in the middle game. Alphago is using a neural network in a new way and it is a big jump for Go programs. Alphago started with human games. If it stopped there than it would play equal to human players. The iterative learning made it stronger than any human player. People have been steadily improving shogi programs for 40 years without a neural network. Brute force is relative term. Technically brute force means look at every move. Nobody can do that in shogi. Not even in chess. Instead, they smartly prune away inferior moves so the programs look deeper than the human player. Alpha beta pruning in Go does not get you a very strong program. Are there professionals chu shogi players? I found a list of publications about computer shogi at chessprogramming.wikispaces.com/Shogi . --Mschribr (talk) 10:49, 25 May 2016 (UTC)Reply

Alphazero

edit

Not sure where the best place for the following content added about Alphazero by Quantanew is, but it certainly does not belong in the lead section. I'm also not sure if it should even be mentioned in the lead section per WP:UNDUE after a suitable place has been found for it later in the article.

In December 2017, Alphazero beats AlphaGo Zero and other top chess and Shōgi programs after only 24 hours of play. After 4 hours of games Alphazero acquired a superhuman level.[1][2]

References

  1. ^ "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm".
  2. ^ "Entire human chess knowledge learned and surpassed by DeepMind's AlphaZero in four hours".

Given the current layout of the article, it doesn't seem to fit anywhere other than perhaps in the "Milestones" section, but technically I'm sure if this would be considered a milestone in computer shogi. Maybe a new section titled "Recent advances" or something similar should be created for this kind of thing, but "recent" seems to be problematic per WP:REALTIME and may require constant updating or it will quickly be dated and I don't before a separate section titled "Alphazero" would be appropriate per WP:UNDUE. Anyone have any suggestions? -- Marchjuly (talk) 05:17, 7 December 2017 (UTC)Reply

In a title Alphazero.Quantanew (talk) 05:18, 7 December 2017 (UTC)Reply
I don't think a stand-alone section is warranted for this per WP:UNDUE, especially for one or two sentences and especially when none of the other programs mentioned in the article do not have their own stand-alone sections. If this would be considered equivalent to the other content listed in "Programming tools" than perhaps it can be added there. Or, as I suggested above, maybe a new section could be created which incorporates the content about Alphazero and other similar content. It will also need to be cleaned up a bit since the citation template parameters are in Spanish (??) and templates are poorly formatted, plus the content itself is a bit puffy for English Wikipedia standards.
FWIW, I am not saying this should not be added to the article, just trying to figure out the best place. You're probably going to find the Chess people watching Computer chess removing it from that article's lead there as well per WP:LEAD. The lead should not really introduce new information, but only be a general summary of what comes later in the article. There's lots of content in the article that is just as significant as Alphazeor's accomplishment, but it's not mentioned in the lead because it doesn't really belong there. -- Marchjuly (talk) 05:42, 7 December 2017 (UTC)Reply
@Ish ishwar and Mschribr: Also pinging Ish iswar and Mschribr for input since they appear to have made the most recent major contributions to the article. -- Marchjuly (talk) 05:49, 7 December 2017 (UTC)Reply
i dont know if this belongs in the lead since i'm not really sure what this article is trying to be about. Seems like AlphaZero could be mentioned in the lead if there is brief description of the various approaches to computer shogi. But, the article itself doesnt really go into any detail about computer shogi architecture. So, it's odd to mention a single example of a single approach (connectionist network + iterative learning) but not mention other other examples of the same approach as well as different approaches (such as rules + alpha beta pruned search).
for the actual text inserted, it reads like a headline. It's not clear how strong AlphaZero really is since they used unusual settings for the rules-based computer shogi opponent with different types of computer hardware. AlphaZero clearly did something, but we dont have a proper comparison.
Perhaps this is nitpicking, but elmo is not the top shogi engine. It was probably the top engine in the summer. Now, the top engine seems to be the Apery SDT5 evaluation function with Qhapaq learning reinforcement combined with the YaneuraOu 4.79 search engine. The guy who does informal measurements with thousands of 5-second per move games states that Apery-Qhapaq is 196–263 Elo points (plus/minus 1 error) higher than the elmo of the 2017 World Computer Shogi Championship. (Its not really an official measurement though. There's no official Elo rating of humans or shogi engines in the shogi world.) And, elmo lost to the Tanuki Ponpoko eval function in last month's Den Ou tournament (although Apery actually seems stronger than Tanuki Ponpoko). This doesnt mean that AlphaZero isnt stronger than these other engines, but we dont really know.
Maybe it goes without saying, but the paper hasnt been peer reviewed yet and AlphaZero has not competed in any formal tournaments yet either.
The Deep Mind folks have to respond to the criticisms of shogi (and chess) folks. I dont know what to say really – seems uncertain so far. Of course, Wikipedia can mention that Deep Mind has made some claims. – ishwar  (speak) 08:10, 7 December 2017 (UTC)Reply
I agree that at the moment it should only be mentioned in the milestones. It is indeed a milestone: the result of 90=2-8 is amazing even if Elmo was not configured optimally. I look forward to published games, and entry into further tournaments and exhibition matches! (Excellent work on this article, by the way. Is it a translation from Japanese Wikipedia?) --IanOsgood (talk) 21:53, 8 December 2017 (UTC)Reply
incidentally, a page in english on some shogi engine concerns: http://www.uuunuuun.com/single-post/2017/12/07/Some-concerns-on-the-matching-conditions-between-AlphaZero-and-Shogi-engine
and, a blog post by the YaneuraOu developer (which is the alpha–beta search part of the elmo mentioned in the Deep Mind paper): http://yaneuraou.yaneu.com/2017/12/07/deepmind%E7%A4%BE%E3%81%8C%E3%82%84%E3%81%AD%E3%81%86%E3%82%89%E7%8E%8B%E3%81%AB%E6%B3%A8%E7%9B%AE%E3%81%97%E5%A7%8B%E3%82%81%E3%81%9F%E3%82%88%E3%81%86%E3%81%A7%E3%81%99/ (in Japanese, of course)
seems that this Japanese discussion is mostly happening in twitter and blog spaces.... – ishwar  (speak) 21:26, 9 December 2017 (UTC)Reply