DeepMind's AlphaZero Game Changer 2 | AlphaZero vs Stockfish 8

[@PowerPlayChess]

If you like my videos do *subscribe* bit.ly/powerplaysubscription and do checkout the *supporting* *options* through Patreon: bit.ly/patreondanielking or through *PayPal* (links in the description)

[@outspan87]

That "77% chance to win" is a little misleading. It's actually an expected score. AlphaZero rates a position with a number from 0-100, with 50 being a draw. Even a score of 74 is closer to a draw (50) than a win (100), so that number doesn't exactly translate to a chance of winning. AZ would evaluate a completely dead draw as 50 but that doesn't mean there's a 50% chance of winning.

[@PowerPlayChess]

Thanks for the clarification! However, I still think that is a pretty astonishing evaluation of its chances considering the position on the board. As I mentioned in the video, this endgame has to be compared with other options earlier which might have seemed promising, but it reckoned this gave it the most chance of winning.

[@chrisiver8506]

So it considered its self slightly better there interesting, I thought it meant it would win that position 77% of the time against its self

[@guest_informant]

@@bobdole7292 Yes. That was from one one of the Sadler videos wasn't it? Either in the video or in a comment.

[@gregtechno506]

Correct, Sadler explains it in at least one of his AZ vs Stockfish analysis video.

[@TheOneMonk]

bob dole rubbish

[@alexismarcq1514]

Alpha Zero showing us that Chess is far from over.

[@petergreen5337]

Precisely

[@enginekid88]

These games are incredible.

[@petergreen5337]

Indeed

[@leonvoulgaris322]

It's strangely satisfying how stockfish is learning how it is to be bullied

[@BboyDschafar]

Everyone that ever lost to a machine must feel so satisfied that the neuronal networks strike back.

[@BboyDschafar]

I even enjoy watching leela chess zero a bit more, because it feels like a pupil of yours is playing the game. You can contribute your hardware to the network and she learns from what you offer.

[@AnandSivaram22]

Plot twist -it was Magnus with a Fritz 5 giving stockfish the beatdown and the smackdown...disguised as A0

[@melrakan]

Are you going to cover one of A0's 6 losses too? I think those would be very interesting as well!

[@melrakan]

@Thinking Z I don't know but it will be fascinating!

[@calebdonovick4596]

The few losses I looked at they were not particularly interesting. They generally consisted of A0 over pressing (repeatedly turning down draws by repetition). Then eventually losing in an end game after 130ish moves

[@giorgosplatias4227]

The Stormfather what's the difference? It learnt chess in four hours! what if it had 10 hours? ever thought of that?

[@forasago]

Alphazero did not lose. Forced opening moves A0 would never play lost.

[@geertj123321]

These game are amazing, thanks Daniel

[@kolektivmozak238]

Yep, Great job Daniel commenting this strange game. I am finding all games by Alpha Zero odd.

[@proflaxis6968]

Its about time someone filleted that fish. Lovely to see The Sweeney back, Matthew and Natasha. I imagine the Game Changer book will be a best seller. It seems we humans need to up our game.

[@thomasmasuch2068]

Thanks Daniel, great job. Do you also have access to the games Alpha Zero lost?

[@julioandresgomez3201]

It seemed to be playing all calm and tranquil, then ¡¡¡BAM!!! g4

[@martymcfly1982]

Daniel the KING

[@evzalxheladini1205]

I can't believe that stockfish lost this ending a pawn up! Doesn't look real!

[@paulgoogol2652]

then you should learn more about endgames. opposite color bishop games are either very decisive or impossible to win just depending on who has the dominating pieces. Also dominating the only open file and keeping 7th rank pressure by itself is worth a pawn. best stockfish could hope for was to play bd7 and trade rooks which would be at least easily drawing for white.

[@aljehine]

Keep going with this games and thank you Daniel .

[@gerstensaft2936]

It will be interesting to see how this data will change human style of playing the game. Who can adapt this activity element? We'll see. :)

[@furkankarakaya26]

I will :)

[@seeingsights]

Great job on your videos!

[@pawn4129]

So in a sense we're all Alpha Zeros when we blunder a pawn! ;)

[@testthewest123]

...if only we could follow up to get a great position and win...

[@peroperic1080]

Good one! We just don't go about it as lightly as AZ, though.

[@pawn4129]

@@testthewest123 maybe in another life :D

[@patrickbertrand6366]

Another great video Danny! :-) I ask myself how close AlphaZero is playing near the maximum skill strenght of chess, if there is any end possible

[@peroperic1080]

Mindblowing endgame. Vow!

[@SimonFoxGuitar]

The endgame problems with this Monte Carlo tree search are really the key thing that needs to be sorted out. We see Leela time and again give up a promising middle game by simplifying into an endgame where perpetual queen checks ruin the day, or some kind of fortress. Maybe A0 is different in some way, but I'm yet to see how. Unless they're willing to put A0 into the main chess engine tournaments, can we assume it will get kicked around like Leela in these endgames? TCEC and Chess.com demonstrated this problem ad-nauseum. It's just not possible to have a neural network recognize forced perpetual check lines that go 30+ moves deep, because the MCTS algorithm isn't probing deep enough. Anyway, that's my area of concern for this research. Possibly a hybrid approach is best. I can't see how A0 can be better than an alpha-beta engine in those endgame positions where we see Stockfish for example probe very deeply along key lines of play. A0 on the other hand is left to hope the neural network can decide whether a line looks like a perpetual, or if there's "hopefully" a way to squeeze out of it and execute the win.

[@SimonFoxGuitar]

The rest of the A0 play is of course amazing, as is Leela. Loving these videos

[@gabepjurs8650]

Daniel, in a January video from "The New Yorker" Kasparov said that 2 of his 4 most memorable games were game 24 of the Karpov match 1985 and 1987 respectively. I was eager to see your analysis on them but found neither in your channel. Would you like to analyze them?

[@PowerPlayChess]

That's a great idea. I love those games. When I find some free time I'll definitely look into that.

[@guest_informant]

These games are mind-bending but in the end despite being a sample of 1000 it's also a sample of 1 with 1000 variations isn't it? A few random thoughts: 1 Re piece activity over material. Isn't it hard to generalise from these games? It's interesting that AZ can outplay SF in this manner, but maybe an AI with a materialistic bent would do just as well. Or better. 2 (related) AZ is _A_ Truth not _The_ Truth. Maybe a different AI would have a different approach to chess. 3 What about the SF victories? 4 Are there any blind spots for AZ? There used to be a favourite material-down, slow-motion attack that worked against early engines bc of the horizon effect I think. Are there, say Mates in 100 that a human could spot, but AZ can't? 5 That was from reading Tim Krabbe's site years ago, which reminds me of some of the tablebase Mates in 500 where the machine would play Perfect Chess in, something like a NNR v NNN ending. How does AZ's play compare with tablebase perfection? 6. Back to 2. Chess.com covered the earlier release of games and they had similarities to Game 1 covered yesterday. Wasn't there an outrageous Qh1 manoeuvre in one game. Because of the stylistic similarity across games it looks (from the available evidence) a lot like AZ is repeatedly "exploiting" a (relative) weakness in SF. My 2 pennorth. And Thanks.

[@ShamaiOfer]

Hi Daniel, can you please add the games moves in this series.

[@fred324gigo]

👍👍👍👍👍 Thank you very much!

[@thecolorsapp1403]

At 11.10, is Kh6 not possible, threatening to trap the rook on g8?

[@BboyDschafar]

Keep em coming. I love it.

[@DavidSchilter]

Great video, Danny! Seems like lots of youtubers are covering these games, and your coverage is tops! And with respect to your coment "...does prevent white from breaking with a6 himself": I guess AlphaZero is male?

[@creationstation4285]

The level of depth here is just ridiculous. A human being wouldn't stand a chance against Stockfish but Alpha Zero just demolishes it!!!

[@RD-un5bh]

ty

[@moonboy5851]

AlphaZero’s endgame is soul destroying.

[@prisonerofwarhammer3814]

Ok, so next time I'm down a pawn in a rook-and-piece endgame I'll be thinking that I'm actually better :P

[@4idior1]

If there were enough time between moves, would magnus c. be able to win against AZ?

[@Socrates...]

The end of dogma, clear our collective chess palette and try new stuff

[@Hellstaff01]

In these games the c8 bishop often seems to be a bad piece.

[@guest_informant]

Yes. This is true in a lot of human games as well, and something that I don't recall seeing in any general opening book - other than, say, problem Bishops in the QGD or the French. If the Bishop is out of the game, then generally so is the Rook on a8. It's worth quite a lot to paralyse these two pieces.

[@warnaoh]

@@guest_informantwell as a grunfeld and KID player I disagree a little about your opinion of this wonderful Bishop ^^

[@TheRubsi]

As somebody who is not an expert on engines and not particularly great at chess ... why is it even so astounding that Alphazero went into this endgame? Sure it looks silly because it is a pawn down and stuff, but even i could tell that white has very active pieces and the initiative. I'd never have found the right moves in a thousand years, but shouldn't at least stickfish know aswell that it is losing by this position? I'd like to know what stockfish says if you put the position on the board. If it thinks it's losing then it should not be surprising that alpha thought the same. If stockfish think it's equal or good for black then all we really see is how bad stockfish actually is compared to an engine that finds the very best moves.

[@rafaelq.8506]

just a lowly human speaking here, but i think stockfish could draw this if it had played like a human and been afraid of the position, rather than say, "fuck it -- i can draw this."

[@iconicon5642]

Alphazero had it "in mind" - probably not though careful what you wish for.

[@Vinvin5114]

PGN please !

[@ShamaiOfer]

Why does 42...a6 cripples the queen-side pawns? (11:00)

[@JanCarlComposer]

prevents pawns forever from coming forward, but that was necessary IMO

[@Roboviewer]

9.c3 is not new. Equinox played it against Stockfish at TCEC in 2014.

[@BboyDschafar]

when did Stockfish realize that it is in deep trouble?

[@emilioarroyomohamed]

What the hell did just happened on that board?

[@kenspencer9895]

77% chance of winning?? I would have put it around 50% max. Are you going to show an AlphaZero loss at some point?

[@kashishmaheshwari6007]

Alpha zero didn't lose

[@chrisiver8506]

I think they said A0 lost 7 times out of a 1000 but I don't know if they released the games

[@kashishmaheshwari6007]

@@chrisiver8506 yeah my mistake. Thanks for clarification

[@maelstrom57]

They had a separate match where they were made to use opening books from the world chess engine championship. A0 lost many games as black. It just doesn't perform as well as Stockfish when forced to play a Sicilian, a French or a CK.

[@chrisiver8506]

@@maelstrom57 whats ck

[@betabenja]

I think 77% chance means it is only really 2% closer to winning than it is to drawing.

[@AnandSivaram22]

Even that is sufficient to evaluate it +/- under normal nomenclature

[@nicknamenescio]

Well, 2 draws and 2 wins out of 4 games (what 77% chance roughly means if it is meant in terms of expected value for the score) is not too bad (same result is 3 wins and 1 loss). On the other hand, if it is supposed to mean in this concrete usage here that White expects to win 77% percent of games from this position, then the expected value score would be even higher because there could still be draws in the remaining 23% of results.

[@YevgeniyShcherbakov]

After AZ calculated 77% chance of win, did it see the ending position where it wins? If so, the game becomes pointless.

[@Achilles94627]

8:10 - "Magnus likes to get these kinds of positions, where he's only playing for two results: either a win or a draw". What's the alternative? Play for a loss? ;)

[@Phurngirathaana]

That's a phrase often used when one side has a slight but lasting advantage

[@chrisiver8506]

Some great players, Tal for example, would play into a position where they could easily lose the game. Magnus doesn't play positions where he can lose very often

[@animestid]

That means that you aren't taking risks at all, it's a position where you are playing for an strategic advantage where your opponent does not have winning chances

[@ex59neo53]

I'd like to watch AZ with blacks :)

[@Narrowcros]

Keep in mind this was only stockfish 8, even during that time stockfish 10 was out.

[@guepardiez]

No. This was at the beginning of 2018, shortly before the release of Stockfish 9.

[@isabelschwab-adamidis5851]

what am I watching here.. i mean under what circumstances did Stockfish play ?! Not only did it not win with a solid position a pawn up in a 2v0 central pawn situation by trading all the pieces and shuffeling RC8 RD8 in the endgame along other passing moves but it lost the position. Quite the piece of work there i guess it got bored of drawing and flipped a coin between winning and loosing...bad luck

[@brianbailey3374]

Engine games are uninteresting to me, as Fischer said its just memorization. Apart from this, I like the rest of your content, more human games please.

[@Sough]

Gee so nice to Not release engine to the public Or entering it in TCEC, just making a PR paper every year. Thanks for nothing, Google