|Jan-05-19|| ||keypusher: Part I of II
Of all the A0 games, one of my favorites.
Daniel King https://www.youtube.com/watch?v=V86...
This is played with a TCEC opening -- the first eight moves are book, but rare, obviously.
click for larger view
A0 characteristically ignores the threat to its e-pawn and continues 9.c3.
(With colors reversed, SF played 9.exf6 and a <scintillating> 231-move draw resulted...the first 45 or so moves were pretty interesting, anyway. Stockfish vs AlphaZero, 2018)
Equally characteristically, SF takes the offer, 9....fxe5 10.Nxe5 Nxe5 11.dxe5 Qxe5 12.Re1 Qc7.
click for larger view
Now things get weird. 13.Bh5+ Ng6 (13....g6 14.Bd4 Rg8 15.Be5 and Bf3 is obviously bad for Black, and 13....Bg6 14.Bg5 also seems inferior to the text move) 14.g4!!? (looks horrible, doesn't it? In a way, the entire rest of the game is a campaign against SF's light-square bishop) 14....Bd3 15.Nb3 (forcing the B to the queenside) Bc4 16.Nd4! (preventing the B from trading itself off for the knight, and also striking at the center) 16....e5 17.b3! Ba6 (not 17....ed 18.Bf4+, of course) 18.Bf4!
click for larger view
A plethora of pins.
Now 19.Ne6 can be answered with 19....Nxf4! 20.Nxc7 Kxc7. Black has only two minor pieces for the queen, but a very imposing array in the center after 21.g5 (to give the B an escape route) 21...Bd6.
White also has what seems like a simple and clear continuation, 19.Bxg6 hg 20.Bxe5, but his kingside weaknesses are apparent after 20....Qd7. Rather than play a complicated middlegame with even material, A0 prefers to head for an endgame a pawn down.
19.Bg3! Re8 (Ne6 was now a real threat) 20.a4 Kb8(?) (my SF 10 prefers 20....Bd6 immediately, and putting the king in the corner, though rather A0-like, turns out to be a bad idea) 21.b4 (continuing the campaign against the light-squared bishop) 21....Bd6 (finally threatening to capture the knight) 22.Nf5 Bc5 23.Qd2 Ka8 24.Qg5! Re6 25.Nxd6 Rxd6 27.Bxg6! (capturing now forces Black to recapture with the rook, keeping the h-file closed) Rxg6 28.Qxe5 Qxe5 29.Rxe5 Rxg4
click for larger view
Remarkably, A0's expected score in this seemingly dry position with five pawns against six was 77% (meaning that it thought it had a better chance of winning than drawing). SF10 also sees White as slightly better.
|Jan-05-19|| ||keypusher: Part II of II
29.f3 Rg6 30.a5! Rf6 31.Rae1 Rff8 (thanks to White's 30th move, 31....a6 32.Re8+ is mate in four) 32.Kf2 g6(?) (somewhere around here, King suggests giving back the pawn with ...d5-d4, allowing the bishop to sit on d5) 33.Re7 h5 34.Be5! Rg8 35.Bg7 (forcing both Black rooks to passive squares) Rc8 36.Bd4 Bb5 37.Kg3 Bd3 38.Kh4 Bf5 39.Kg5 Rcd8 40.R1e2 Rc8 (incidentally, SF10 keeps wanting to play ...a6, which seems like a good idea, especially once you get to move 43) 41.R1e2 Rcd8 42.h4 Rc8
click for larger view
43.a6! +- with the idea of giving Black weaknesses on both sides of the board that the bishop will be unable to cover. Black could have stopped this with 42....Bd3, but then 43.R2e6 Bf5 44.Rf6 Rde8 45.Rxe8 Rxe8 46.f4! setting up Rxg6! and White's kingside pawns will overpower the rook.
Obviously hopeless, but 43....b6 44.Rb7 Rce8 45.Rxe8 Rxe8 46.Rc7 Bd3 47.Rxc6 Bxa6 48.Rxg6 and again Black will lose on the kingside.
44.Bd4 Kb8 (forced, since Black can't allow White to double rooks on the 7th file. But now Black is completely passive.)
45.Bxa7+ Ka8 46.Bc5 Kb8 47.Ra2 (resuming the campaign against the bishop) Bd3 48.Rd2 Bb1 49.Rd1 Bc2 50.Ba7+ Ka7 51.Rde1 Bd3 52.Bb6 Kb8 53.Bd4 Bf5 54.f4 Rcd8 55.Ra1 Bd3 56.Re3
click for larger view
Again forcing the bishop to the queenside.
56....Bb5 57.Rae1 Kc7 58.Kh6 Kc8 59.Re6
And now g6 cannot be defended -- the culmination of the campaign against the bishop.
59....Ba4 60.Bc5! (more precise than 60.Rxg6 Rxg6 61.Kxg6 Rf8 62.f5 Bc2!, though White is still winning after 63.Bf6) 60....Kb7 61.Rxg6 Rxg6+ 62.Kxg6 d4(?)
At this point SF starts flinging away pawns, no doubt trying to stop the game from being published. Sorry Stockfish!
|Jan-05-19|| ||keypusher: <22.Nf5 Bc5 >
Should be 22....Bc4, of course.
|Jan-05-19|| ||SChesshevsky: I'm not sure what happened but I would think even SF would suppose that losing something like 6 tempo for the e-pawn and getting a position like that after 19...Re8 wasn't a very good deal.|
Wondering if other good engines end up there as well?
|Jan-05-19|| ||keypusher: <SChesshevsky: I'm not sure what happened but I would think even SF would suppose that losing something like 6 tempo for the e-pawn and getting a position like that after 19...Re8 wasn't a very good deal.>|
Huh? Six tempo? Two, by my count, And, Dr. Tarrasch taught us you need 3 tempi to justify sacrificing one pawn. :-)
At move 9 my SF10 plays like SF8 did in this match, and its top line runs 9....fe 10.Nxe5 Nxe5 11.de Qxe5 12.Re1 Qc7 13.Bh5+ Ng6 14.g4 Bd3 15.Nf3(?), with a 0.00 evaluation at 41 ply.
You seem to think that SF8 played badly in these games. I don't think that is true.
|Jan-05-19|| ||SChesshevsky: <keypusher> I get five direct tempo. The queen backup, knight, bishop. Then b3 bishop push and 0-0-0. But then you get pin city which is probably gonna cost more.|
At 19 ...Re8 it looks like the CG computer says .00 but blacks a pawn up so the position is closer to 1.00. Pretty bad if not lost probably. I'd think it would be hard to find many who would want black at that point.
Pretty bad or busted by move 20 doesn't seem that impressive.
Maybe it was the forced ...f6? I wonder how common is that and what's the holding percentage?
|Jan-05-19|| ||keypusher: <SChessevsky>
White gains two tempi between Blackís 9th move and Whiteís 13th ó the rook goes to e1 and itís whiteís move instead of blackís. Neither side gains any tempi after that, unless you count pawn moves, which neither Tarrasch nor I would do, particularly when they are moves like b2-b3 and g2-g4!?.
A0 launches its gambit with the modest 9.c2-c3, definitely not a tempo gainer, and the key move in the game is 19.Bf4-g3, a clear loss of tempo.
The ďbookĒ position after 8 moves is evidently quite rare (there are no examples in the cg database). Black is behind in development and is nevertheless opening the center, which is dubious according to classical precepts. In other words, I think the truly questionable moves in the opening were played before SF took over. Incidentally, I think second-rate lines like this were why some of the A0-SF games played with TCEC openings were blowouts, relatively speaking (viz. Stockfish vs AlphaZero, 2018 and AlphaZero vs Stockfish, 2018) ó the engines are forced to play positions they would never choose on their own. But we get some fascinating games that way, so Iím not complaining.
|Jan-05-19|| ||keypusher: keypusher>
<At 19 ...Re8 it looks like the CG computer says .00 but blacks a pawn up so the position is closer to 1.00. Pretty bad if not lost probably. I'd think it would be hard to find many who would want Black at that point.>
Again, huh? The computer says 0.00 (so does my SF10), but Blackís up a pawn, so itís ďreallyĒ +1 for White? Iím sorry, do pawns not count for you?
And yes, Iíd be happy to play Black in that position. Iíve got a big pawn center and Iíve solved my development problems, Iím a pawn up, and my opponent has a gaping self-inflicted wound in his kingside and his bishop is out of play on h5. Iím going to gain another tempo when I play ...Bd6 forcing his N to move, and if he regains the pawn with Bxg6 and Bxe5 his kingside weaknesses are going to be glaring. If I didnít know better from having played through the game a million times, Iíd be very surprised by the 0.00 evaluation.
|Jan-05-19|| ||AylerKupp: <<SChesshevsky> At 19 ...Re8 it looks like the CG computer says .00 but blacks a pawn up so the position is closer to 1.00. Pretty bad if not lost probably.>|
Computer evaluations are much more than just counting material. Considerations such as pawn structure, development, king safety, space and square control, threats, and many more are taking into account, suitably weighted as to their relative importance. So it's very possible that the computer evaluation of positional factors can be enough to compensate for any material disadvantage, and it's obvious that a forced mate in one with that player to move is enough to compensate for any positional disadvantage.
Whether you agree or not with the computer's evaluation of the position being equal is a different issue, of course. But an evaluation disadvantage of [1.00] means that the opponent has a (barely) significantly better position, not necessarily a win. And, as <keypusher> pointed out, a 3 tempi advantage in the opening is considered to be approximately equal to a pawn, so having an extra pawn with a 5 or 6 tempi disadvantage an evaluation of [0.00] is not unreasonable.
|Jan-05-19|| ||SChesshevsky: <AylerKupp> I'm not sure what the stats are but I'm wondering how many computer v. computer games with one having an eval of 1.00 are saved. Probably very few if any is my guess.|
If the eval is 0.0 a pawn up then it would only make sense that if one drops the pawn with no compensation then the eval would likely be worse for the former holder of that pawn. If you drop the pawn with no comp and are also still positionally weak the eval probably goes against you pretty good.
This happens a couple of times in this game. SF ends up material and after awhile that material is gone and the eval just gets worse and worse. After 43..bxa6, isn't SF up two pawns but it doesn't take long for the eval to really turn against him.
This scenario plays out not just in this game but others as well. I just get the feeling, that for whatever reason these SF games feel more like 1998 computer chess than 2018 computer chess.
In this match SF seems to take material, even material that might be helpful or material that seriously weakens his position. Then the eval looks equal because the extra material counts as something but really is positionally very bad with no counterplay. So it starts out 0.0, then goes like +.33 then close to 1.0. All without SF seemingly being in the game.
Striking is the difference in counterplay A0 got in the same opening game referenced above.
I just find it strange that a standard top 2018 engine would frequently get itself into such passive positions when it's opponent plays generally just natural moves.
|Jan-05-19|| ||SChesshevsky: Related to my last post, here's an interesting example where SF as black is 3 pawns up at 32...Bd7 and the CG computer has the eval at +1.07.|
AlphaZero vs Stockfish, 2018
|Jan-05-19|| ||SChesshevsky: Hey, another one:
At 42...Nb7, SF as black is 2 pawns up. CG eval is 0.0. At 50...c4, eval at +.50. At 55...Rf6 eval at +1.47.
AlphaZero vs Stockfish, 2018
|Jan-05-19|| ||keypusher: <If the eval is 0.0 a pawn up then it would only make sense that if one drops the pawn with no compensation then the eval would likely be worse for the former holder of that pawn. If you drop the pawn with no comp and are also still positionally weak the eval probably goes against you pretty good.>|
White could have won the queen at move 19 and gotten a negative evaluation (19.Ne6 Nxf4! 20.Nxc7 Kxc7) or regained the pawn (19.Bxg6 hg 20.Bxe5 Qd7) with a 0.30 evaluation from my SF at 40 ply. So it's not like the 0.00 evaluation after 19....Re8 (which my SF maintains after an hour, incidentally) is secretly terrible for Black but masked by an extra pawn.
<SChesshevsky: Related to my last post, here's an interesting example where SF as black is 3 pawns up at 32...Bd7 and the CG computer has the eval at +1.07.>
Seriously, SChesshevsky? Here's the position, white to move.
click for larger view
A modern engine, a 1998 engine, and your average club player could tell that Black has got serious problems. Do you think SF8 was thrilled with how things had gone? I'm sure its evaluation was at least as bad as my SF10's evaluation's is. You seem to think SF8 <wanted> positions like this. No one and no thing does!
At that point SF had been at least two pawns up for 16 moves. It had refused a third pawn at move 18 (which it probably should have taken, incidentally). It had taken a pawn at move 27 and then immediately given it back. It had no decent alternative to taking a third pawn at move 30 -- judge for yourself. It was offered a fourth pawn on move 32, which it refused. And it had been hit with a whole series of remarkable moves by its opponent.
You seem to think SF sleepwalked through these games picking up every pawn on offer. That's not what happened.
(Incidentally, my SF10 evaluates the position at move 16 (when it's two pawns up) at 0.00 at 41 ply. It plays the opening exactly the same way SF8 did in the match.)
<SChesshevsky: Hey, another one:
At 42...Nb7, SF as black is 2 pawns up. CG eval is 0.0. At 50...c4, eval at +.50. At 55...Rf6 eval at +1.47.>
This is just more of the same. Here's the position at move 42.
click for larger view
I'm surprised the cg website evaluation is only 0.00, frankly. My SF eval is +0.32 at 39 ply. The position at move 50 is similar. And I hope you noticed that at move 55 SF is the exchange down, with more material about to go.
In that game, SF had been a pawn up since move 4 (not by its choice, obviously) and had taken a second pawn on move 41, by which time its position was already looking questionable to this patzer, and would have remained questionable whether the pawn was taken or not. So whatever got SF into trouble in this game, it wasn't blind greed.
Finally, you're drawing conclusions from an unrepresentative sample of games. SF and A0 played 1000 games without opening books. (I don't know the exact breakdown for the other games, except that SF was more likely to win games played with pre-set openings.) A0 won 155 games, SF won six, and the rest were drawn -- nearly 85% of the total. So obviously games in which SF went way up in material but wound up losing were exceptions, even if all A0's wins were like the game in the first diagram above, which is not the case.
One other point about this game. As I noted earlier, I'd be very happy to play Black at move 19, if I didn't know better from experience. If you told me I was going to lose, and asked me to guess how it would come about, you know what I definitely <wouldn't> predict? A double-rook-and-opposite-color bishop ending where I had an extra pawn. That (along with 14.g4, and the sustained anticlerical campaign) is what blows my mind about this game.
Bottom line -- questions of hardware and configuration aside, AlphaZero is playing amazing chess. You seem determined not to credit what is right before your eyes.
|Jan-06-19|| ||keypusher: <SChesshevsky>|
One more for you.
click for larger view
Black's a piece up, but horribly undeveloped and obviously going down in flames. In fact it resigned here.
So what's different about this game? AlphaZero was Black (as I'm sure you know).
Whatever else it might have been, this match was not the chessic equivalent of Goofus vs. Gallant.
|Jan-06-19|| ||perfidious: <keypusher>, bit surprising really that Alpha spit the bit without playing on a while, but of course there is no good defence to Rxh7 and devastation to follow on those wretchedly weak dark squares.|
|Jan-06-19|| ||SChesshevsky: <keypusher> It's pretty clear that you're a big A0 fan. I also agree that it plays excellent chess.|
My take is that it's tough to tell how good as it's SF that is playing sub optimally. In my opinion.
You mentioned at some point SF didn't want a certain position but that's what it went for. And in those circumstances the lines didn't seem forced. In this game it was SF who set the p!ay with taking the pawn ending ...Qxe5.
So the analysis really doesn't have anything to do with A0 but where SF went wrong and what if any are the improvements. I think it was the e5 grab idea or if that line is best then back at ...f6 that wasn't good.
It usually takes a bad move or idea to lose. So where did SF go wrong and are there any improvements?
|Jan-06-19|| ||keypusher: <You mentioned at some point SF didn't want a certain position but that's what it went for. And in those circumstances the lines didn't seem forced. >|
If we're talking about the position after move 32 in AlphaZero vs Stockfish, 2018, I don't think it makes sense to talk in terms of SF <going for> that position, any more than A0 <went for> that position I posted from its blowout loss to SF. Or I <go for> the terrible positions I generally wind up in against stronger opponents.
<So where did SF go wrong [in this game] and are there any improvements?>
Putting aside 8....f6, which seems to be dubious, SF's clearest error seems to be 20....Kb8 (+0.67 at 40 ply) instead of ...Qd8 (+0.19 at 41 ply) or ...Bd6. In qualitative terms, 20....Kb8 seems to give A0 a more or less straight path to the favorable ending arising at move 28, where Black's weak back rank is one of White's biggest assets.
Later on, SF might have given back the pawn with ...d5-d4, giving its bishop a more secure perch (Daniel King's suggestion). And it should have played ...a7-a6 at some point rather than just sitting tight and sliding its rook back and forth between d8 and c8.
In human terms, taking the pawn at move 9 was risky, but probably objectively best. I have no doubt 9....fxe5 would get the best engine evaluation for Black at any reasonable search depth.
I let my SF10 run at move 9. After the first few seconds, its top choice is unvaryingly 9....fxe5. At 10 minutes/40 ply, its evaluation is +0.13, i.e. it thinks White is slightly better. It's certainly not thinking, "Hey! Free pawn!"
Its line runs 9....fxe5 10.Nxe5 Nxe5 11.dxe5 Qxe5 12.Re1 Qc7 13.Qb3 e5 14.Nf3 Be4 15.Rad1. It hasn't seen (or isn't impressed by) the Bh5+/g2-g4 idea.
At 42 ply/16 minutes, the eval has risen to +0.44, but it's still playing 13.Qb3 e5 14.Nf3 Be4 15.Rad1.
At 43 ply/27 minutes, the eval is +0.45 and the line is 13.Qb3 c5 14.Qb5+ Nc6 15.Bxc5.
Even going forward to White's 13th move, SF10's top line at 40 ply/10 minutes is 13.Qb3 c5 14.Qb5+. Maybe that's objectively better than 13.Bh5+, who knows. Regardless, that reinforces my sense that 13.Bh5+ Ng6 14.g4 was a really remarkable conception from A0.
|Jan-06-19|| ||SChesshevsky: <keypusher> Thanks for the excellent analysis! Related to SF getting itself into bad positions, that's what is confusing. Seems these engines can look out at least 15 to 20 moves for hundreds or more lines. At least at TCEC computer play. So I'd guess that SF likely saw the possibility of the poor resulting position 15 to 20 moves ahead but still evaluated the position as playable enough to go for it. Just seems strange, that in many match games, SF with that kind of move foresight could get such positionally really bad positions knowing it was a possibility 20 moves ahead of time.|
I agree that A0 13.Bh5+ is very good but certainly seems logical. Gains a tempo and forces some sort of weakness. I wonder why your SF is not impressed? Maybe black eventual ...c5 freeing move which can be something looked for in CK advanced?
I also agree that the ...Qc7 ...f6 opening idea is probably dubious. Making it very tough to equalize in my opinion. I haven't checked out the match that closely to see for any other CK variations. But if they are using book, hopefully they tested out many more mainlines Might be more useful, at least for me.
|Jan-06-19|| ||AylerKupp: <<SChesshevsky> I'm not sure what the stats are but I'm wondering how many computer v. computer games with one having an eval of 1.00 are saved. Probably very few if any is my guess.>|
I don't know either. But consider that if the evaluation was [0.01] less (an insignificant amount) then it would be subjectively assessed as "White stands definitely better". Would that change your guess?
Don't get me wrong, guesses are fine. But unless you're a top-level GM (and even then), they would still be just guesses although in the latter case they would be educated guesses. But still guesses. I just prefer to have some data to back up whatever conclusion I "guess" at.
<After 43..bxa6, isn't SF up two pawns but it doesn't take long for the eval to really turn against him.>
That's what evals often do in complex positions, particularly if the engine generating the evaluations is subject to the horizon effect; they can swing widely. Remember that AlphaZero's evaluations are based on simulated games conducted to the end and a statistical assessment of the results. So, I don't think, AlphaZero is subject to the horizon effect, at least not as much as "classical" engines like Stockfish. And, running on much more powerful hardware, it can get more accurate assessment of the most likely results of each move.
<I just find it strange that a standard top 2018 engine would frequently get itself into such passive positions when it's opponent plays generally just natural moves.>
I do too. Maybe Stockfish simply didn't have enough time to calculate properly. But that's just a "guess" on my part. :-)
|Jan-07-19|| ||keypusher: <I agree that A0 13.Bh5+ is very good but certainly seems logical. Gains a tempo and forces some sort of weakness. I wonder why your SF is not impressed? Maybe black eventual ...c5 freeing move which can be something looked for in CK advanced?>|
Oh, I agree 13.Bh5+ is quite rational ó itís following it up with g2-g4 that is surprising. Like Daniel King, if I saw a beginner doing that Iíd explain to him what a terrible idea it was.