chessgames.com
Members · Prefs · Laboratory · Collections · Openings · Endgames · Sacrifices · History · Search Kibitzing · Kibitzer's Café · Chessforums · Tournament Index · Players · Kibitzing

(If you register a free account you won't see all these ads!)
AlphaZero (Computer) vs Stockfish (Computer)
AlphaZero - Stockfish (2017), London ENG, Dec-04
Queen's Indian Defense: Classical Variation. Polugayevsky Gambit (E17)  ·  1-0
ANALYSIS [x]

FEN COPIED

explore this opening
find similar games 9 more AlphaZero/Stockfish games
sac: 24.Qd6 PGN: download | view | print Help: general | java-troubleshooting

TIP: Some games have photographs. These are denoted in the game list with the icon.

PGN Viewer:  What is this?
For help with this chess viewer, please see the Olga Chess Viewer Quickstart Guide.
PREMIUM MEMBERS CAN REQUEST COMPUTER ANALYSIS [more info]

A COMPUTER ANNOTATED SCORE OF THIS GAME IS AVAILABLE.  [CLICK HERE]

Kibitzer's Corner
< Earlier Kibitzing  · PAGE 2 OF 2 ·  Later Kibitzing>
Dec-10-17
Premium Chessgames Member
  chessgames.com: We stuck Stockfish on 21.? for 5 hours to see if it could come up with 21.Bg5!!

The answer is "no" analysis of move 21.?

Dec-10-17
Premium Chessgames Member
  zanzibar: <<chessgames> We stuck Stockfish on 21.? for 5 hours to see if it could come up with 21.Bg5!!

The answer is "no" analysis of move 21.?>

Really?!

I got 21.Bg5 after just 19-ply from my version of Stockfish:

<

<-0.1 Depth: 19 Nodes: 3334K (889 kn/s) Time: <3.75 s[!]>>

<<1 19 -0.10 21.Bg5>> f5 22.Qf3 d5 23.Nf6 gxf6 24.Bxf6+ Rxf6 25.Re8+ Qg8 26.Rxg8+ Kxg8 27.Re1 Nd7 28.Re7 Nbc5 29.b4 Re6 30.Rxe6 Nxe6 31.Qxf5 Ndf8 32.Qf6 Rc8 33.Qxh6 Be2 34.f4 Ng7 35.f5 Nxh5 (3.75)

<2 19 -0.18 21.b4> d5 22.Nc3 Bc4 23.Bf4 Qg8 24.b5 Na5 25.Bd6 Re8 26.bxc6 Nbxc6 27.Nxd5 Rxe1+ 28.Rxe1 Re8 29.Rxe8 Qxe8 30.Ne3 Qe6 31.Qxe6 Bxe6 32.a4 (3.75)

<3 18 -0.26 21.Nc3> d5 22.b4 Nd6 23.Nxd5 cxd5 24.Bxd5 Bb7 25.Qf3 Bxd5 26.Qxd5 Nd7 27.Qxd6 Rad8 28.Qf4 Nf6 29.Qf3 Rd3 30.Be3 Kg8 31.Rac1 Nd5 32.b5 (3.75)

<<>>>

Disbelievers can find a screenshot here, along with some cryptic commentary:

https://zanchess.wordpress.com/2017...

PS- Damn it <CG>, when are we going to get the round numbers put in for these games, so I can easily refer to this specific game (or any other for that matter)!?

And putting the round number in the game heading on the kibitz list would allow us to keep track of which game is getting what comment, btw.

.

Dec-10-17
Premium Chessgames Member
  WorstPlayerEver: <zanzibar>

I belieb the line 21. Bg5 f5 22. Qf3 d5 23. Nxf6 gxf6 24. Bxf6+ Rxf6 25. Re8+ Qg8 26. Rxg8+ Kxg8 27. Qc7 is stronger for White

Dec-10-17
Premium Chessgames Member
  zanzibar: <WPE> I was evaluating the position from after Black's move 20.

If you play out the game, and then let the engine churn, like <CG> did, 21.Bg5 doesn't show up as a leading candidate move (for me, I typically set MPV = 3, and so it didn't show up in the top-3).

But, and here's the point, play 21.Bg5 (and maybe a few more ply), then back up to the position after 20...Kh8. Now the engine quickly finds 21.Bg5 as leading candidate move (again, one always has to give the depth, i.e. 19-ply).

This is rather odd, that it seems after 5 hours the engine didn't look at the branch of 21.Bg5 to see it was good, but if you force it to (via a walk-thru), it does.

.

Dec-10-17
Premium Chessgames Member
  zanzibar: For the record - I never trust the engine moves to far beyond the first candidate moves in a "projection".

For best lines I would suggest a "walk-thru" - say, allowing the engine to go to a fixed depth, play the leading candidate, then rinse and repeat, etc. etc.

Dec-11-17
Premium Chessgames Member
  Zhbugnoimt: 19..Qe6 loses to 20.g4! (Houdini 5 depth 28)
Dec-13-17  jdoucette: Stockfish 8 sees 21. Bg5 on a measly 8-core machine (not 64-core they were using) in only 1 hour 9 minute, with it obviously winning at 1 hour 22 minutes.

AlphaZero had 1,500 times the computing power of Stockfish (4 TPUs with 180 TFlops *each*, whereas a 72-core monster PC CPU can barely do 0.5 TFlops). But Stockfish only needed 10 times more power to see the same move. What would it do with another 150 times more power on top of that, to put it at equal footing as AlphaZero?

(Or to put it another way, reduce AlphaZero to the same hardware as Stockfish, and 80,000 moves/sec is suddenly only 5 moves/sec.)

Dec-14-17  devere: <chessgames.com: We stuck Stockfish on 21.? for 5 hours to see if it could come up with 21.Bg5!! The answer is "no">

Stockfish 8 64 BMI2 saw 21.Bg5 at depth 47 if not sooner (I left it running all night on my puny 5th generation i3 laptop). It's mostly about computing power, and AZ heavily outgunned SF. It wasn't designed to be a fair fight, just an impressive publicity stunt for Google.

Dec-14-17
Premium Chessgames Member
  WorstPlayerEver: It's interesting if you really know how to combine all the knowledge involved. It could also be very dangerous, though.

Let's say I could build an AI robot, and it has the ability to clone itself.

With simple commands I would rule and regulate you within 4 days. If you are lucky :)

Dec-18-17
Premium Chessgames Member
  chessgames.com: <Stockfish 8 64 BMI2 saw 21.Bg5 at depth 47 if not sooner> We only went to depth 43, so somewhere between 44 and 47 it must find the move. (Ironically it finds the move in the first 1-2 seconds of computation and then quickly discards it.)

We should run the position again for a really long think just for fun.

Dec-18-17
Premium Chessgames Member
  AylerKupp: <chessgames.com>, <zanzibar> With regards to getting different results when you each ran an analysis of the position after 20...Kh8, you may both be right. Remember that multi-core chess engines, at least "classical" ones using alpha-beta pruning, are notoriously non-deterministic. If each of you ran an analysis of the position after 20...Kh8 on the same computer, with the same engine, to the same search depth, you will get different results. Not <MAY>, <WILL>. Guaranteed. The results might be slight changes in the evaluations or major changes in the evaluations resulting in different move rankings. Thus, using MPV=3, <chessgames.com>'s Stockfish might evaluate 21.Bg5 as its top move yet <zanzibar>'s Stockfish (in obviously a different computer) might not evaluate 21.Bf5 as one of its top 3 moves. Repeat your analyses and you will see.

I ran a very, very limited test (i.e. one position) using Stockfish 8 and MPV=5 as described starting in US Championship (2017) (kibitz #624) of the position after 13...Bxd4 from W So vs Onischuk, 2017.


click for larger view

I later ran similar tests using Stockfish 8 with 1, 2, 3, and 4 threads and this is a summary of the evaluation results. Much easier to compare if you copy the text to MS Word and replace all the blanks with tabs so that the numbers line up.

Average Evaluation by Move, d=35 for all analyses:

Move 1 thread 2 threads 3 threads 4 threads
14.d3 [+0.25] [+0.28] [+0.34] [+0.36]
14.f4 [+0.32] [+0.23] [+0.32] [+0.35]
14.Qd5 [+0.29] [+0.33] [+0.32] [+0.29]
14.Qa6 [+0.19] [+0.26] [+0.28] [+0.29]
14.Qc6 [+0.25] [+0.18] [+0.19] [+0.24]
14.Qb4 -- [+0.08] -- --

Rank by Average Evaluation, d=35 for all analyses:

Move 1 thread 2 threads 3 threads 4 threads
14.d3 3 2 1 1
14.f4 1 4 3 2
14.Qd5 2 1 2 --
14.Qa6 5 3 4 3
14.Qc6 3 5 5 4
14.Qb4 -- 6 -- --

I've run similar tests with Houdini 4 and Komodo 10 and have seen similar results. Similar results have also been indicated by other users from different positions (naturally), most notably <RandomVisitor> using Rybka 4.1.

Dec-18-17
Premium Chessgames Member
  AylerKupp: <zanzibar> For best lines I would suggest a "walk-thru" - say, allowing the engine to go to a fixed depth, play the leading candidate, then rinse and repeat, etc. etc.>

What you are describing is similar to a process commonly referred to as "forward sliding", and this has been around for quite a while. See, for example, "Modern Chess Analysis", 2004, by Robin Smith. But I would caution you not to just allow the engine to go to a fixed (and presumably relatively small) search depth; let it run for an adequate search depth ("adequate" is chess engine dependent), move forward a few moves in its principal variation, and then restart the analysis.

The reason is the horizon effect. It's well known that the evaluations by an engine based on the last move it analyzed in its search tree is pretty much worthless since the engine cannot "see" beyond that move. The most extreme example is a position I once analyzed when the engine indicated that White has a somewhat advantageous position even though Black has a mate in 1 on its next move. And here is an example of where Komodo 9.2 (the strongest engine at that time) came up with a favorable evaluation for White even though Black had a forced mate in 5: Stockfish (Computer) (kibitz #115).

But the horizon effect can be more subtle than that. If you use an engine, say, Stockfish, to analyze a position to d=35, the first few moves in its principal variation have the benefit of a deep search and you can therefore can be fairly confident in its evaluation But the last few moves only have the benefit of a shallow search and you should not be very confident in its evaluations. Therefore, since the evaluations from the search tree's leaf nodes could be suspect, you need to be careful to attach a great amount of significance to the evaluation of <ANY> computer line, regardless of its depth.

To try to detect this situation I actually use your approach and I call it "forward leaping", an obvious play on "forward sliding". The objectives are different; in forward sliding one is trying to find the best move by gradually pushing forward the consequences of the horizon effect. But this obviously can take a lot of time. So in forward leaping all I'm trying to do is determine whether the engine's initial evaluation is reasonable; if after restarting the analysis at the end position the resulting evaluation is similar to the initial evaluation (they will seldom, if ever, be the same), then I think I can reasonable confidence in the initial evaluation. But if the evaluations are greatly different, then you have to resort to either forward sliding or backward sliding (another process) to determine where the engine went wrong in its initial evaluation.

And I should say that several users at this site have expressed doubts about the usefulness of "forward leaping", but at least you're proof that great minds do think alike. ;-)

Dec-18-17
Premium Chessgames Member
  AylerKupp: <chessgames.com> We only went to depth 43, so somewhere between 44 and 47 it must find the move. ... We should run the position again for a really long think just for fun.>

Sure, I'm glad that you have lots of time on your hands. :-) But you should keep in mind what I have egotistically and tongue-in-cheek have identified as "AylerKupp's Corollary to Murphy's Law" (AKC2ML):

"If you use your engine to analyze a position to a search depth=N, your opponent's killer move (the move that will refute your entire analysis) will be found at search depth=N+1, regardless of the value you choose for N."

Dec-19-17  StupidQuestion: Would love to see AlphaZero play the Sicilian.
Dec-28-17  SirChrislov: It really downgrades the public's view of the match once you have knowledge of the fairness and connect it.

A 10-pack of entertaining BS is what my mind made up of it.

Jan-08-18  sankukay50: Rebel 3.0 (2000) recommends 19...Qf5 after 15m running
Jan-08-18
Premium Chessgames Member
  RandomVisitor: <chessgames.com: We stuck Stockfish on 21.? for 5 hours to see if it could come up with 21.Bg5!! The answer is "no" analysis of move 21.?>

I stuck the current development build of Stockfish: Stockfish_18010422_x64_modern on the same position:

Depth time score move

43 0:33:34 +0.88 21.b4 d5 22.Nc3...
44 0:36:04 +0.61 21.b4 d5 22.Nc3...
45 0:43:33 +0.79 21.b4 d5 22.Nc3...
46 1:04:15 +0.81 21.b4 d5 22.Nc3...
47 1:20:23 +0.88 21.Bg5 <first hint of 21.Bg5, not a full pv>

47 1:47:01 +1.12 21.Bg5 f5 22.Qf4... <first full PV with 21.Bg5>

8 cores 4.00 GHz AMD FX-8370 32GB system

Jan-08-18
Premium Chessgames Member
  whiteshark: REQUEST ANALYSIS


click for larger view

White to move

1) +0.81 (32 ply) 22.Qf4 d5 23.Be7 dxe4 24.Bxf8 Nd7 25.Ba3 Ndc5 26.Re3 Nd3 27.Qc7 Qg8 28.Qxc6 Ne5 29.Qc7 Ng4 30.Ree1 Qe8 31.f3 Nf6 32.fxe4 fxe4 33.Qf4 Qxh5 34.Qh4 Qf7 35.Bxe4 Nxe4 36.Rxe4 Rd8 37.Re7 Qf5 38.Qf4 Qxf4 39.gxf4 Rd2 40.Rae1

5.0 minute analysis by Stockfish 8 v270317

So after <21.Bg5> SF understands the seriousness of the situation quite well.

Jan-08-18
Premium Chessgames Member
  WorstPlayerEver: ★sighs★
Jan-08-18
Premium Chessgames Member
  RandomVisitor: After 16.Nc3 white really has nothing for the two pawns. It seems that Black can simply return them for a drawn game.


click for larger view

Stockfish_18010422_x64_modern:

<48/67 37:31 +0.08 16...Ne8> 17.Bf4 Bc4 18.Bxb8 Rxb8 19.Qxd7 c5 20.Rad1 a5 21.Re4 Be6 22.Qa7 Rd8 23.Qxb6 Rxd1+ 24.Nxd1 Bg4 25.Qxf6 Nxf6 26.Re1 Be6 27.Nc3 Rd8 28.a3 Rd4 29.Rd1 g5 30.Bf1 Rxd1 31.Nxd1 Kf8 32.Ne3 Ke7 33.Kg2 Ng4 34.Nxg4 Bxg4 35.f4 gxf4 36.gxf4 Be6 37.Kf3 Bd5+ 38.Kg4 Kf6 39.Bd3 Be6+ 40.Kf3 Bf5 41.Be2 Ke6 42.Ke3 Kd5 43.b3 Be6 44.Bc4+ Kd6 45.Ke4 Bxc4 46.bxc4 a4

Jan-09-18
Premium Chessgames Member
  RandomVisitor: <Wikipedia> In AlphaZero's chess tournament against Stockfish 8 (2016 TCEC world champion), <each program was given one minute's worth of thinking time per move>. Stockfish was allocated 64 threads and a hash size of 1 GB, a setting that Stockfish's Tord Romstad later criticized as suboptimal.

AlphaZero was trained on chess for a total of nine hours before the tournament. During the tournament, AlphaZero ran on a single machine with four application-specific TPUs. In 100 games from the normal start position AlphaZero won 25 games as white, won 3 as black, and drew the remaining 72.

In a series of twelve 100-game matches (of unspecified time or resource constraints) against Stockfish starting from the 12 most popular human openings, AlphaZero won 290, drew 886 and lost 24.

Jan-09-18
Premium Chessgames Member
  RandomVisitor: After 16...Nb7? it is almost game over


click for larger view

Stockfish_18010422_x64_modern:

<58/93 15:38:35 +0.96 17.Ne4 Qg6 18.h4 h5> 19.Bf3 d5 20.Bxh5 Be2 21.Qxe2 Qxe4 22.Qxe4 dxe4 23.Rxe4 c5 24.Re7 Na5 25.Be3 Nbc6 26.Rc7 Rae8 27.Bg5 g6 28.Bf3 Rc8 29.Rxc8 Rxc8 30.Bf6 Nd4 31.Bxd4 cxd4 32.Rd1 Rd8 33.b3 Rd6 34.Kf1 Nc6 35.Bxc6 Rxc6 36.Ke2 Rd6 37.Kd3 Kg7 38.Rc1 b5 39.Rc7 Ra6 40.Rc2 Rd6 41.Re2 Kf6 42.f3 Kg7 43.Rh2 Rf6 44.Rf2 Rd6 45.g4 a6 46.Rc2 Rf6 47.Ke4 d3 48.Rc1 Rd6 49.Rd1 Rd8 50.Rxd3 Re8+ 51.Kd4 Rh8 52.Kc5 Rxh4 53.Kb6 b4 54.Kxa6

Feb-18-18
Premium Chessgames Member
  chessgames.com: <RV> <I stuck the current development build of Stockfish: Stockfish_18010422_x64_modern on the same position:

Depth time score move

43 0:33:34 +0.88 21.b4 d5 22.Nc3...
44 0:36:04 +0.61 21.b4 d5 22.Nc3...
45 0:43:33 +0.79 21.b4 d5 22.Nc3...
46 1:04:15 +0.81 21.b4 d5 22.Nc3...
47 1:20:23 +0.88 21.Bg5 <first hint of 21.Bg5, not a full pv>

47 1:47:01 +1.12 21.Bg5 f5 22.Qf4... <first full PV with 21.Bg5>

8 cores 4.00 GHz AMD FX-8370 32GB system>

We did a 15 hour probe last night, with the new Stockfish 9, and got it up to 53 plies. It turns out that it finds b4 (which it regards as pretty good), then it finds 19.Bg5 around 47/48 plies just like you found, but then after an even deeper probe it switches back to b4 again. However once the move 16.Bg5!! is made it recognizes that White is better off than if 19.b4 was played. (You can see this analysis just by clicking on a position and engaging the "engine" button in Olga.)

So SF is still deficient in its grasp of the position (at least, so it seems) but at least the centipawn differential is not enough to make it actually criticize 16.Bg5 as inferior.

Feb-18-18  sudoplatov: One method of mitigating the Horizon Effect is to use a quiescence search. One only evaluates positions with no "active" moves (captures, checks, promotions, etc.) and follows all capture and check chains to quiescence. Using a hash table helps here by extending the depths of searches. (
Feb-19-18
Premium Chessgames Member
  WorstPlayerEver: I still have to check with SF9, but I have some closed positions where SF8 keeps on moving back and fro. While I regard the positions as winning for White.

Because the positions are around move 20, it takes a lot of research though.

search thread:   
< Earlier Kibitzing  · PAGE 2 OF 2 ·  Later Kibitzing>
NOTE: You need to pick a username and password to post a reply. Getting your account takes less than a minute, totally anonymous, and 100% free--plus, it entitles you to features otherwise unavailable. Pick your username now and join the chessgames community!
If you already have an account, you should login now.
Please observe our posting guidelines:
  1. No obscene, racist, sexist, or profane language.
  2. No spamming, advertising, or duplicating posts.
  3. No personal attacks against other members.
  4. Nothing in violation of United States law.
  5. No posting personal information of members.
Blow the Whistle See something that violates our rules? Blow the whistle and inform an administrator.


NOTE: Keep all discussion on the topic of this page. This forum is for this specific game and nothing else. If you want to discuss chess in general, or this site, you might try the Kibitzer's Café.
Messages posted by Chessgames members do not necessarily represent the views of Chessgames.com, its employees, or sponsors.
Spot an error? Please submit a correction slip and help us eliminate database mistakes!
This game is type: CLASSICAL (Disagree? Please submit a correction slip.)

Featured in the Following Game Collections [what is this?]
2017 games
by Severin
21 B-N5 .... give me a minute
from tea4twonty's favorite games by tea4twonty


home | about | login | logout | F.A.Q. | your profile | preferences | Premium Membership | Kibitzer's Café | Biographer's Bistro | new kibitzing | chessforums | Tournament Index | Player Directory | Notable Games | World Chess Championships | Opening Explorer | Guess the Move | Game Collections | ChessBookie Game | Chessgames Challenge | Store | privacy notice | contact us
Copyright 2001-2018, Chessgames Services LLC