|Jan-04-19|| ||Sargon: If anyone knows of any other games from this 2018 AlphaZero - Stockfish "Match" that are not in our database—but available in PGN format—please provide a link.|
|Jan-04-19|| ||Sargon: Some of these games were basically started from "book" positions, as indicated by the erroneous annotation symbols. I'll be sure to clean things up, and we'll probably ultimately separate this current batch of 210 games into two discrete tournaments.|
For the moment, it's easy enough to tell which games fit which category.
|Jan-11-19|| ||jith1207: Thanks for doing this.|
|Jan-12-19|| ||HeMateMe: It's nice that so many games were produced, it wasn't always like this.|
|Jan-12-19|| ||alexmagnus: I see many draws ending after 256 moves. Was there a weird "256 move rule" in place?|
|Jan-12-19|| ||keypusher: <alexmagnus> Yes, they'd call it a draw after 512 "half-moves," mercifully.|
|Jan-12-19|| ||HeMateMe: Why did stock win the first eight decisive games but then lose the match?|
|Jan-12-19|| ||keypusher: What makes you think these games are in order?
They played a 1000-game match without opening books, which A0 won +155 -6 =839. IIRC 110 of those games have been released and are in the CG database (including all six SF wins). They also played other ways, including I-don’t-know-how-many games using preselected openings taken from the 2016 engine championships. One hundred of those games have been published and are in the database. You can identify them by the pencil icon in list view or the “book” notation in the game score.
Further details here. https://deepmind.com/blog/alphazero...
|Jul-09-19|| ||csmath: Keep in mind that this so called "match" is bogus. Stockfish is not represented by anybody (open source) and therefore you can do whatever disadvantage you want with SF to promote your own. |
A0 depends on "networks" and SF does not, installation of Stockfish is simple while A0 is one ridiculous nonsense when you need to install it.
On independent web site like CEGT that has no reason to promote either Stockfish is still ahead of A0.
I installed A0 on my computer and I returned to SF, that is my opinion. It is not worth effort bothering.
Once the authors of A0 make an effort to create engine that is independent of network libraries I'll take a second look, for now they are not worth the hype they produced.
|Jul-09-19|| ||keypusher: <csmath>
<I installed A0 on my computer>
Oh no you didn't. Unless you're Google's cg account. In which case, those bastards should pony up for a premium membership.
|Jul-09-19|| ||csmath: Leela Chess.|
|Jul-18-19|| ||Sally Simpson: ***
Posting here instead of Kaissa vs Chaos, 1974 (kibitz #7) where I last posted.
Have printed out the game names and page numbers and pasted them in the back of the book. (thanks again)
Had to use two pages.
No problem, I gleefully pasted one over the advert for NIC.
No game index, names NOT appearing where they should be in the name index......And the same game with different names and notes!
This game Stockfish vs AlphaZero, 2018
page 138 'take my Knight' where are told after 5...Nh6 we are already in an original and unexplored position.
Onto page 278 and we see the same game only this time it is called the 'not so quiet game.' and we leave established theory by move 7. (the note is a moot point - but same game with a different name threw me.)
I'm starting to get it.
Same Game - different name = 'Game Changer.'
|Jul-18-19|| ||keypusher: <No game index, names NOT appearing where they should be in the name index......And the same game with different names and notes!>|
Yes, I think the same thing happened with AlphaZero vs Stockfish, 2018. Not really ceegee's fault, but you can't fit two of Sadler's titles in the space provided for a game note in a collection ...sometimes you can't even fit one. I should have put in an asterisk or something when he named a game more than once.
Which is really a terrible thing for Sadler to have done, I must say. One obvious takeaway is that if they had done an index, they surely would have noticed the error and corrected it.
|Jul-19-19|| ||Sally Simpson: ***
Under normal circumstances a minor nitpick error but with no index and such a thick book a bit frustrating.
The same game Stockfish vs AlphaZero, 2018 also has the strange note on 280 when discussing move 40.
The note says S.F. gave back the exchange to blunt White's attack.
S.F. has actually been the exchange down since move 24. (probably meant extra pawns.)
Enjoying the games and spotting the wee crafty two move traps Alpha sets, some are blatant but this game Stockfish vs AlphaZero, 2018 is loaded with them and Mathew (I think anyway) misses a good chance to give a variation rather than that very long spiel on page 159 after 39.Nb2
and how ineffective being in command of the only file is to White.
I'm going to the game, nobody has noted it up so far. I'll build my review around it, but want to post there before someone else does.
|Jul-19-19|| ||Sally Simpson: ***
Forgot to add (see above)
Been reading about the future plans for Alpha, there are going to put it onto solving global warming.
I hope not. I know what it's solution will be - get rid of the humans.
|Jul-19-19|| ||AylerKupp: <<csmath> Keep in mind that this so called "match" is bogus. Stockfish is not represented by anybody (open source) and therefore you can do whatever disadvantage you want with SF to promote your own. >|
Yes you can. And in the first 100-game match they did exactly that. Some of them were a possibly misguided attempt to make the match seem more equal; since AlphaZero did not support opening books or endgame tablebases and did not have a time management function they did not configure Stockfish to use the first two and made the time control a fixed one minute per move.
DeepMind was also somewhat sloppy in specifying the hardware configuration used by Stockfish. They indicated that Stockfish used 64 threads but they did no specify whether they used 64 cores or 32 cores with hyperthreading, a big difference. They were also silent on the type of server used, its speed, and memory. And they also had a relatively small hash table for Stockfish (1 GB) given that it was executing with 64 threads.
Possibly as a result of the criticism they received, things were a lot better on the second series of matches in early 2018. They specified that they used the TCEC 44 core using two 2.2 GHz Intel Xeon CPUs with 22 cores each, a 32 GB hash table, and Syzygy tablebases, and a time control of 3 hours/game with 15 sec increments per move. Still no opening book for the games starting from the initial position.
But, frankly, I think that whatever tweaking DeepMind might have done to Stockfish did not have a significant effect on the results of the matches compared to my estimated 80X computational performance advantage enjoyed by AlphaZero. For example, I ran a 10-game match between a "full" Stockfish 8 (StockFull) and a "crippled" Stockfish 8 (StockCrip) with a time control of 1 minute/more. StockFull used 4 threads, a 128 MB has table, Perfect_2010.abk opening book, and 5-piece Syzygy tablebase support. I was limited to a 5-piece version of the Syzygy tablebases since the 6-piece tablebase probing code does not work on my archaic 32-bit system. StockCrip used 1 thread, a 16 MB hash table (analogous to the hash table used in the late 2017 AlphaZero vs. Stockfish 8 match; 1 GB hash table / 64 threads = 16 MB hash table/thread), no opening book and no tablebase support.
StockFull won the match with a score of +2, =7, -1. Ten games are hardly enough to indicate definitive results but with this limited sample of games there is no evidence that StockCrip performed significantly worse than StockFull. This seems to indicate that Stockfish 8 as configured for the AlphaZero match was not significantly impacted by its less than optimum configuration.
So I still refer to the results of the various matches to determine which is the better engine as, at best, "inconclusive".
|Jul-19-19|| ||AylerKupp: <<Sally Simpson> I hope not. I know what it's solution will be - get rid of the humans.>|
I think that you've been watching "Avengers – Infinity War" and applied it's "logic".
It does remind me of a story I read of an attempt to use AI on operating systems. A user was beta testing a version of the system and was trying to save a large file. It got a series of messages along the following lines:
> Insufficient disk space to save the file.
> Please wait
> File saved.
The OS had deleted half of the user's files, at random, in order to make enough room to save the file that the user was trying to save.
|Jul-19-19|| ||MrMelad: Hi <csmath>, I failed to understand how your inability to install Leela chess has anything to do with the bogusness of the AlphaZero vs Stockfish match??|