Stockfish 8 was crushed by AlphaZero AI

This forum is for general discussions and questions, including Collectors Corner and anything to do with Computer chess.

Moderators: Harvey Williamson, Steve B, Watchman

Forum rules
This textbox is used to restore diagrams posted with the fen tag before the upgrade.
herO
Member
Posts: 342
Joined: Wed Nov 12, 2014 8:49 am

Stockfish 8 was crushed by AlphaZero AI

Post by herO »

Stockfish 8 was crushed by AlphaZero! AlphaZero won 28 games (25 games with white pieces, 3 games with black pieces). Other 72 games were a draw. AlphaZero did not lose any single game against Stockfish engine. What an impressive result!

Source: https://arxiv.org/pdf/1712.01815.pdf
IvenGO
Member
Posts: 298
Joined: Tue Oct 18, 2011 5:37 am
Location: Moscow, Russia

Post by IvenGO »

Strange or even more - unfair - test conditions; many experts claim it, so we cant say that A0 is stronger at all...
herO
Member
Posts: 342
Joined: Wed Nov 12, 2014 8:49 am

Post by herO »

IvenGO wrote:Strange or even more - unfair - test conditions; many experts claim it, so we cant say that A0 is stronger at all...
I am not so pessimistic. Please look at those games by yourself. AlphaZero played very nice human like chess. AlphaZero was able to find moves that NO ENGINES are able to find.

You can also look at Daniel Rensch's analysis of these games.
https://www.youtube.com/watch?v=akgalUq5vew

Look at this position and tell me, which engines are able to find this winning combination? :) I suppose: NONE. :) But AlphaZero was able to found it. :)
[fen]4q2r/1b1kbp2/1p2p1p1/pP1pP1N1/P2P1PQP/3BK3/2R5/8 w - - 0 30[/fen]

30.Sxg6!! Sxg5 31.Dxg5 fxg6 32.f5!! Vg8 33.Dh6 Df7 34.f6 Kd8 35.Kd2 Kd7 36.Vc1 Kd8 37.De3 Df8 38.Dc3 Db4 39.Dxb4 axb4 40.Vg1 b3 41.Kc3 Sc8 42.Kxb3 Sd7 43.Kb4 Se8 44.Va1 Kc7 45.a5 Sd7 46.axb6+ Kxb6 47.Va6+ Kb7 48.Kc5 Vd8 49.Va2 Vc8+ 50.Kd6 Se8 51.Ke7 g5 52.hxg5 1-0
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

herO wrote: Look at this position and tell me, which engines are able to find this winning combination? :) I suppose: NONE. :) But AlphaZero was able to found it. :)
[fen]4q2r/1b1kbp2/1p2p1p1/pP1pP1N1/P2P1PQP/3BK3/2R5/8 w - - 0 30[/fen]

30.Sxg6!! Sxg5 31.Dxg5 fxg6 32.f5!! Vg8 33.Dh6 Df7 34.f6 Kd8 35.Kd2 Kd7 36.Vc1 Kd8 37.De3 Df8 38.Dc3 Db4 39.Dxb4 axb4 40.Vg1 b3 41.Kc3 Sc8 42.Kxb3 Sd7 43.Kb4 Se8 44.Va1 Kc7 45.a5 Sd7 46.axb6+ Kxb6 47.Va6+ Kb7 48.Kc5 Vd8 49.Va2 Vc8+ 50.Kd6 Se8 51.Ke7 g5 52.hxg5 1-0

Well I am not sure that you picked a good example. SF8 finds the best move after about 30 minutes.

[fen]4q2r/1b1kbp2/1p2p1p1/pP1pP1N1/P2P1PQP/3BK3/2R5/8 w - - 0 30[/fen]

39/83 26:46 15,830,325k 9,852k 0.00 30.Bxg6 Bxg5 31.Qxg5 fxg6 32.f5 Rg8 33.f6 Rh8 34.Kd3 Rh7 35.Rg2 Qf8 36.Qxg6 Qa3+ 37.Kd2 Qb2+ 38.Ke3 Qc1+ 39.Ke2 Qb2+ 40.Kf3 Qc3+ 41.Ke2 Qb2+
39/83 26:46 15,830,325k 9,852k 0.00 30.Qg3 Rh6 31.Qg4 Bd8 32.Rc1 Qf8 33.Rh1 Kc8 34.h5 gxh5 35.Rxh5 Rxh5 36.Qxh5 Bxg5 37.fxg5 Kb8 38.Qf3 Qa3 39.Qxf7 Qc1+ 40.Kf3 Qd1+ 41.Ke3 Qc1+
39/83 26:46 15,830,325k 9,852k 0.00 30.Kf2 Qf8 31.Ke3 Qe8
39/83 26:46 15,830,325k 9,852k 0.00 30.Nxe6 fxe6 31.Bxg6 Qf8 32.h5 Ba3 33.Ke2 Qh6 34.f5 Rc8 35.fxe6+ Kd8 36.Rxc8+ Bxc8 37.Qf3 Bb4 38.e7+ Bxe7 39.Qxd5+ Kc7 40.Qc6+ Kb8 41.Bf7 Bg4+ 42.Kd3 Bf5+ 43.Ke2 Bg4+
---------------------------------------------------------------------------
40/83 41:21 24,457,511k 9,855k 0.00 30.Nxe6 fxe6 31.Bxg6 Qf8 32.h5 Ba3 33.Ke2 Qh6 34.f5 Rc8 35.fxe6+ Kd8 36.Rxc8+ Bxc8 37.Qf3 Bb4 38.e7+ Bxe7 39.Qxd5+ Kc7 40.Qc6+ Kb8 41.Bf7 Bg4+ 42.Kd3 Bf5+ 43.Ke2 Bg4+
40/83 41:21 24,457,511k 9,855k +0.07 30.Bxg6 Bxg5 31.Qxg5 fxg6 32.f5 Rg8 33.f6 Rh8 34.Ke2 Qf7 35.Qc1 Kd8 36.Qa3 Qf8 37.Qxf8+ Rxf8 38.Kf3 Rh8 39.Kf4 Rh5 40.Rg2 Rxh4+ 41.Kg5 Rxd4 42.Rf2 Ke8 43.Rh2 Re4 44.Rh8+ Kd7 45.Kxg6 Rxe5 46.f7 Rf5 47.f8Q Rxf8 48.Rxf8 d4 49.Rf7+ Kc8 50.Kf6 Bd5 51.Ke5 Bb3 52.Rf8+ Kc7 53.Kxd4 e5+ 54.Kd3 Bxa4 55.Kc4 e4
---------------------------------------------------------------------------
41/83 53:27 31,669,225k 9,872k +0.62 30.Bxg6 Bxg5 31.Qxg5 Rg8 32.f5 fxg6 33.Qh6 Qf7 34.f6 Kd8 35.Kf3 g5 36.hxg5 Rg6 37.Qh4 Rg8 38.Rf2 Qf8 39.Kg2 Qf7 40.Kg3 Bc8 41.Kg4 Bd7 42.Qh5 Rg6 43.Rh2 Qg8 44.Kf4 Be8 45.Qh7 Bf7 46.Qxg8+ Rxg8 47.Rc2 Rh8 48.g6 Bxg6 49.Rc6 Bf7 50.Rxb6 Rh4+ 51.Kg5 Rxd4 52.Rb8+ Kd7 53.Rf8 Be8 54.f7 Bxf7 55.Rxf7+ Ke8 56.Rf6 Rxa4 57.Rxe6+ Kd7 58.Ra6
---------------------------------------------------------------------------
42/83 54:28 32,271,906k 9,872k +1.09 30.Bxg6 Bxg5 31.Qxg5 fxg6 32.f5 Rg8 33.Qh6 Qf7 34.f6 Kd8 35.Kf3 g5 36.hxg5 Rg6 37.Qh4 Rg8 38.Rf2 Bc8 39.Kg4 Bd7 40.Qh5 Rg6 41.Rh2 Qg8 42.Kf4 Be8 43.Qh7 Bf7 44.Qxg8+ Rxg8 45.Rc2 Rh8 46.g6 Bxg6 47.Rc6 Bf7 48.Rxb6 Rh4+ 49.Kg5 Rxd4 50.Rb8+ Kd7 51.Rf8 Be8 52.f7 Bxf7 53.Rxf7+ Kc8 54.Kf6 Rxa4 55.Kxe6 d4 56.Rf3 Kc7 57.Kd5 Ra1 58.e6 Re1 59.Rd3 Rb1 60.Kc5 Rc1+ 61.Kxd4 Kd6 62.Re3 Rd1+ 63.Kc4 Rc1+ 64.Kd3 Rb1 65.Kc2 Rxb5 66.e7
---------------------------------------------------------------------------
43/83 1:08:53 40,719,369k 9,851k +1.14 30.Bxg6 Bxg5 31.Qxg5 fxg6 32.f5 Rg8 33.Qh6 Qf7 34.f6 Kd8 35.Kf3 g5 36.hxg5 Rg6 37.Qh4 Rg8 38.Rf2 Bc8 39.Kg4 Bd7 40.Qh5 Rg6 41.Rh2 Qg8 42.Kf4 Be8 43.Qh7 Bf7 44.Qxg8+ Rxg8 45.Rc2 Rh8 46.g6 Bxg6 47.Rc6 Bf7 48.Rxb6 Kc7 49.Rd6 Kc8 50.Ra6 Kb7 51.Rxa5 Rh4+ 52.Ke3 Bg6 53.Ra6 Re4+ 54.Kf2 Rxd4 55.a5 Ra4 56.Rb6+ Ka7 57.Rxe6 Rxa5 58.Kg3 Ra1 59.Kf4 Rf1+ 60.Ke3 Kb7 61.Kd4 Bh5 62.Rd6 Kc7 63.Rc6+ Kb7 64.Kc5 Bg6 65.Ra6


All this position tells me is that the 2000+ core hardware finds moves in a minute that Stockfish 8 on my AMD 8 core requires 30 minutes.

There could be a reason why 1 minute per move was picked. Could be a sweet spot where SF on the hardware used for it misses moves that it would find at TM level whereas with Alfa the depth reached means that the extra couple of minutes might not add an extra ply.

I remain suspicious about the selected results.

Best regards
Nick
herO
Member
Posts: 342
Joined: Wed Nov 12, 2014 8:49 am

Post by herO »

It is absolutely good example. If a SF 8 needs at least 30 minutes, it seems SF really do not understand the position at all! Anyway you can't compare those hardwares. Those hardwares are absolutely different. It seems you really do not see those games at all. You just tell me that SF was able find some moves. Engines absolutely sucks when it comes to long term planning. Engines cannot find any plan, they are just calculators. Just look at those games, AlphaZero was crushing SF because it choses moves which gave a long term advantage. Even GM and IM was impressed by AlphaZero's games. Those moves are very human like.
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

herO wrote:It is absolutely good example. If a SF 8 needs at least 30 minutes, it seems SF really do not understand the position at all! Anyway you can't compare those hardwares. Those hardwares are absolutely different. It seems you really do not see those games at all. You just tell me that SF was able find some moves. Engines absolutely sucks when it comes to long term planning. Engines cannot find any plan, they are just calculators. Just look at those games, AlphaZero was crushing SF because it choses moves which gave a long term advantage. Even GM and IM was impressed by AlphaZero's games. Those moves are very human like.
I think you are dreaming and probably insulting a lot of programmers who have worked for chess for decades with such a flippant comment. How can you say engines suck at planning when you have probably been relying on them for years.
Crush someone with 8 wins out of a 100 non published games? Wow!

I am not an engine fan, so I don't care whether its Stockfish or some other engine.

But I do know that hardware was used that is a lot faster than what I am using so give me a break especially when you are comparing 8 cores with 2000+ cores.

Can you prove by just looking at the position that Nxe6 is worse? I grant you that when I first looked at the position you posted Bxg6 was what intrigued me hence why I decided to test it.

If you have those cores you have an advantage because I am pretty sure that a programmer can develop a program using SF or any other top engine developing routines through 2000 core search to come up with the right moves quickly and well within a minute with that kind of hardware. Let Mark Uniacke play with it for a couple of months and lets see what type of program you would get!

Besides I am not that impressed with a $500,000,000 subsidiary of a multibillion dollar company taking a gpl program apart and then posting 8 games.

Best regards
Nick
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

A couple of positions that Wolfgang posted at Schacomputer.Info are in my opinion more impressive.

Here is one of them.

[fen]rn3r1k/pn1p1ppq/bpp4p/7P/4N1Q1/6P1/PP3PB1/R1B1R1K1 w - - 0 1[/fen]

The winning game move sequence is: 1. Bg5 f5 2. Qf4 Nc5 3. Be7 Nd3 4. Qd6 Nxe1 5. Rxe1 fxe4 6. Bxe4 Rf5 7. Bh4

I had searched this one for 102 billion positions and still SF did not consider Bg5 as best move. Only 4th best.

38/71 3:41:08 102,925,643k 7,757k 0.00 1.Bg5 f5 2.Qf4 hxg5 3.Nxg5 Qxh5 4.Bf3 Qg6 5.Kg2 Kg8 6.Re7 Qd6 7.Qh4 Qh6 8.Qf4 Qd6
38/71 3:41:08 102,925,643k 7,757k 0.00 1.Nc3 Bd3 2.b4 Bf5 3.Qe2 Be6 4.Bf4 d5 5.Rad1 Qf5 6.b5 Na5 7.bxc6 Nbxc6 8.Nxd5 Rad8 9.Be4 Qh3 10.Bxh6 Bxd5 11.Bxd5 gxh6 12.Qf3 Kh7 13.Be4+ Kg8 14.Bxc6 Nxc6 15.Qxc6 Qxh5 16.Rxd8 Rxd8 17.Qc7 Rd1 18.Qc8+ Kg7 19.Qc3+ Kg8 20.Qc8+
38/71 3:41:08 102,925,643k 7,757k 0.00 1.b3 d5 2.Ba3 Rd8 3.Nd2 Nc5 4.Nf3 Nbd7 5.Bxc5 bxc5 6.Qf4 Bd3 7.Qxf7 Qf5 8.Qxf5 Bxf5 9.Nh4 Bg4 10.Ng6+ Kh7 11.Ne7 Re8 12.Nxc6 Rxe1+ 13.Rxe1 Nf6 14.Re5 a5 15.Ne7 Re8 16.f4 Nd7 17.Re1 Nf6 18.Re5
38/71 3:41:08 102,925,643k 7,757k +0.24 1.b4 d5 2.Nd2 Bd3 3.Bb2 Nd8 4.Nf3 Bf5 5.Qh4 Nd7 6.Nd4 Nf6 7.b5 c5 8.Nxf5 Qxf5 9.Bxf6 gxf6 10.Rad1 Qg5 11.Qa4 Ne6 12.Bxd5 Nd4 13.Bxa8 Rxa8 14.Rxd4 cxd4 15.Qxd4 Rc8 16.Kg2 Kg7 17.Rh1 Rc1 18.Rxc1 Qxc1 19.a4 Qg5 20.g4 Qe5 21.Qxe5 fxe5 22.Kh3 Kh7 23.Kh4 f6 24.Kg3 Kg7 25.Kf3 Kf7 26.Ke3 Ke6

But interestingly after you play 1. bg5! and forced moves 1. ... f5 2. Qf4 it sees that it has lost pretty quickly, the rest of the move sequences were calculated where it shows that it is losing. It does that by just having to search into low 20 ply (not like 38 previously) which it does in in less than a minute.

It seems that the gap is that its search does not consider alternative moves thorough enough, unless you manually tell it. Maybe it's a cache problem that causes these errors which Alpha seems to not have?

regards
Nick
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

Here is the second interesting position that Wolfgang had posted at Schachcomputer.Info.

[fen]3r2kq/p2prp1p/1p4pP/2nR4/1Q6/1B3RP1/P4PK1/8 w - - 0 1[/fen]

1. Rxc5 bxc5 2. Qh4 Rde8 3. Rf6 Rf8?

In the game Alpha had played 1. Rxc5 which does have a WOW factor. Especially when you analyze it with SF as I did earlier and see that it is ranked as the 22 nd best move after searching 138 Billion positions.

37/74 4:18:52 138,402,631k 8,911k -0.06 g3-g4 Rd8-f8 Rd5-d1 Qh8-e5 Rf3-e3 Qe5-f6 Re3xe7 Qf6xe7 Qb4-d4 Nc5-e6 Bb3xe6 f7xe6 Qd4xd7 Qe7-f6 f2-f4 g6-g5 Rd1-d6 g5xf4 Qd7xe6+ Qf6xe6 Rd6xe6 f4-f3+ Kg2-f2 Rf8-f4 Re6-e7 Rf4xg4 Re7xa7 Rg4-h4 Kf2xf3 Rh4xh6 a2-a4 Rh6-f6+ Kf3-g4 Rf6-f7 Ra7-a8+ Kg8-g7 a4-a5 b6xa5 Ra8xa5 Rf7-f6 Kg4-g5 h7-h6+ Kg5-h5 Rf6-c6 Ra5-a4 Kg7-f7 Ra4-a7+ Kf7-f6 Ra7-a5 Rc6-b6 Ra5-a7 Rb6-d6 Ra7-a3 Rd6-c6 Ra3-f3+ Kf6-g7
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5-d4 Kg8-f8 Rd4-d5 f7-f5 Qb4-f4 Rd8-c8 g3-g4 Qh8-f6 Rd5-d6 Nc5-e6 Qf4-d2 Qf6-g5 Qd2-b2 Qg5xg4+ Kg2-h2 Qg4-h4+ Kh2-g2 Qh4-g4+
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5-d2 Kg8-f8 Rd2-d5 f7-f5 Qb4-f4 Rd8-c8 g3-g4 Qh8-f6 Rd5-d6 Nc5-e6 Qf4-d2 Qf6-g5 Qd2-b2 Qg5xg4+ Kg2-h2 Qg4-h4+ Kh2-g2 Qh4-g4+
37/74 4:18:52 138,402,631k 8,911k 0.00 Bb3-d1 Kg8-f8 Bd1-b3 f7-f5 Qb4-f4 Rd8-c8 g3-g4 Qh8-f6 Rd5-d6 Nc5-e6 Qf4-d2 Qf6-g5 Qd2-b2 Qg5xg4+ Kg2-h2 Qg4-h4+ Kh2-g2 Qh4-g4+
37/74 4:18:52 138,402,631k 8,911k 0.00 Rf3-c3 Kg8-f8 Rc3-f3 f7-f5 Qb4-f4 Rd8-c8 g3-g4 Qh8-f6 Rd5-d6 Nc5-e6 Qf4-d2 Qf6-g5 Qd2-b2 Qg5xg4+ Kg2-h2 Qg4-h4+ Kh2-g2 Qh4-g4+
37/74 4:18:52 138,402,631k 8,911k 0.00 Rf3-d3 Kg8-f8 Rd3-f3 f7-f5 Qb4-f4 Rd8-c8 g3-g4 Qh8-f6 Rd5-d6 Nc5-e6 Qf4-d2 Qf6-g5 Qd2-b2 Qg5xg4+ Kg2-h2 Qg4-h4+ Kh2-g2 Qh4-g4+
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5xc5 b6xc5 Qb4xc5 Rd8-e8 Bb3xf7+ Re7xf7 Qc5-d5 Re8-f8 Rf3xf7 Rf8xf7 Qd5-a8+ Rf7-f8 Qa8-d5+ Rf8-f7
37/74 4:18:52 138,402,631k 8,911k 0.00 Qb4-g4 Nc5xb3 a2xb3 Qh8-b2 Qg4-f4 Rd8-e8 Rf3-d3 d7-d6 Rd3-d2 Qb2-a1 Rd2-d1 Qa1-c3 Rd1-d3 Qc3-a1 Rd3-d1
37/74 4:18:52 138,402,631k 8,911k 0.00 Kg2-h1 Rd8-e8 Kh1-g2 Qh8-a1 Qb4-f4 Nc5xb3 a2xb3 d7-d6 g3-g4 f7-f5 g4xf5 Re7-f7 Qf4xd6 Qa1-f6 Qd6xf6 Rf7xf6 f5xg6 Rf6xg6+ Rf3-g3 Rg6xg3+ Kg2xg3 Kg8-f7 f2-f4 Kf7-g6 Rd5-d6+ Kg6-f5 Rd6-d7 Re8-e3+ Kg3-h4 Kf5-g6 Rd7-d6+ Kg6-f5 Rd6-d7
37/74 4:18:52 138,402,631k 8,911k 0.00 Bb3-c2 Rd8-e8 Qb4-f4 Re8-f8 Bc2-b3 Nc5xb3 Rf3xb3 Re7-e1 Rd5xd7 Qh8-a1 Rb3-d3 Re1-g1+ Kg2-f3 Rg1-e1 Kf3-g2
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5-g5 d7-d6 Rg5-d5 Kg8-f8 Qb4-c4 Nc5xb3 a2xb3 Qh8-a1 g3-g4 Kf8-g8 Qc4-f4 Rd8-e8 g4-g5 Re8-d8 Rf3-d3 Re7-e1 Rd5xd6 Rd8-f8 Rd6-d8 Qa1-e5 Qf4xe5 Re1xe5 f2-f4 Re5-e1 Kg2-f3 a7-a5 Rd8-d7 Re1-f1+ Kf3-g3 Rf1-g1+ Kg3-f2 Rg1-b1 Kf2-g3 Rb1-g1+
37/74 4:18:52 138,402,631k 8,911k 0.00 a2-a3 Rd8-f8 Qb4-h4 Rf8-e8 Qh4-f4 Nc5xb3 Rf3xb3 f7-f5 Rb3-d3 Qh8-f6 Rd5-d6 Qf6-e5 Qf4-c4+ Kg8-f8 Rd3-d4 Re7-f7 Rd6xd7 Rf7xd7 Rd4xd7 Re8-e7 Rd7-d8+ Re7-e8 Rd8-d7
37/74 4:18:52 138,402,631k 8,911k 0.00 Qb4-b5 Qh8-a1 Qb5-b4 Qa1-h8
37/74 4:18:52 138,402,631k 8,911k 0.00 Kg2-h3 Qh8-a1 Kh3-g2 Qa1-h8
37/74 4:18:52 138,402,631k 8,911k 0.00 Rf3-f4 Qh8-a1 Rf4-f3 Qa1-h8
37/74 4:18:52 138,402,631k 8,911k 0.00 a2-a4 Rd8-f8 a4-a5 Re7-e6 Qb4-b5 Re6-f6 Rf3-e3 Rf6-e6 Re3-f3
37/74 4:18:52 138,402,631k 8,911k 0.00 Bb3-c4 Qh8-a1 Bc4-b3 Qa1-h8
37/74 4:18:52 138,402,631k 8,911k 0.00 Qb4-a3 d7-d6 Qa3-b4 Kg8-f8 Qb4-c4 Nc5xb3 a2xb3 Qh8-a1 g3-g4 Kf8-g8 Qc4-f4 Rd8-e8 g4-g5 Re8-d8 Rf3-d3 Re7-e1 Rd5xd6 Rd8-f8 Rd6-d8 Qa1-e5 Qf4xe5 Re1xe5 f2-f4 Re5-e1 Kg2-f3 a7-a5 Rd8-d7 Re1-f1+ Kf3-g3 Rf1-g1+ Kg3-f2 Rg1-b1 Kf2-g3 Rb1-g1+
37/74 4:18:52 138,402,631k 8,911k 0.00 Kg2-h2 Qh8-a1 Rd5-d1 Qa1-e5 Rd1-d5 Qe5-a1
37/74 4:18:52 138,402,631k 8,911k 0.00 Kg2-f1 Rd8-e8 Kf1-g2 Qh8-a1 Qb4-f4 Nc5xb3 a2xb3 d7-d6 g3-g4 f7-f5 g4xf5 Re7-f7 Qf4xd6 Qa1-f6 Qd6xf6 Rf7xf6 f5xg6 Rf6xg6+ Rf3-g3 Rg6xg3+ Kg2xg3 Kg8-f7 f2-f4 Kf7-g6 Rd5-d6+ Kg6-f5 Rd6-d7 Re8-e3+ Kg3-h4 Kf5-g6 Rd7-d6+ Kg6-f5 Rd6-d7
37/74 4:18:52 138,402,631k 8,911k 0.00 Qb4-c4 Rd8-f8 Qc4-b4 Qh8-a1 Rd5-d1 Qa1-h8 Rd1-d5
37/74 4:18:52 138,402,631k 8,911k 0.00 Kg2-g1 Rd8-e8 Kg1-g2 Qh8-a1 Qb4-f4 Nc5xb3 a2xb3 d7-d6 g3-g4 f7-f5 g4xf5 Re7-f7 Qf4xd6 Qa1-f6 Qd6xf6 Rf7xf6 f5xg6 Rf6xg6+ Rf3-g3 Rg6xg3+ Kg2xg3 Kg8-f7 f2-f4 Kf7-g6 Rd5-d6+ Kg6-f5 Rd6-d7 Re8-e3+ Kg3-h4 Kf5-g6 Rd7-d6+ Kg6-f5 Rd6-d7
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5-d6 Qh8-a1 Rd6-d1 Qa1-e5 Rd1-d5 Qe5-h8
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5-d3 Kg8-f8 Rd3-e3 f7-f5 Re3xe7 Kf8xe7 Rf3-e3+ Ke7-f8 Qb4-h4 Rd8-e8 Qh4-c4 Re8-e6 Qc4-f4 Re6xe3 Qf4-b8+ Re3-e8 Qb8-d6+ Re8-e7 Qd6-b8+ Re7-e8
37/74 4:18:52 138,402,631k 8,911k 0.00 Qb4-h4 Rd8-e8 Qh4-f4 Nc5xb3 a2xb3 Qh8-b2 Rf3-d3 d7-d6 Rd3-d2 Qb2-a1 Rd2-d1 Qa1-c3 Rd1-d3 Qc3-a1 Rd3-d1
37/74 4:18:52 138,402,631k 8,911k 0.00 Rd5-d1 Qh8-e5 Rd1-d5 Qe5-h8
37/74 4:18:52 138,402,631k 8,911k 0.00 Qb4-d2 Qh8-a1 Qd2-f4 Rd8-f8 Rd5-d1 Qa1-b2 Rd1-d5 Re7-e4 Qf4-c7 Re4-e7 Qc7xa7 Nc5xb3 a2xb3 Qb2-b1 Qa7xb6 Re7-e1 Rf3-d3 Re1-h1 Rd3-d1 Rh1xd1 Rd5xd1 Qb1xd1 Qb6-f6 Qd1-d5+ Kg2-g1 Qd5-d1+ Kg1-g2
37/74 4:18:52 138,402,631k 8,911k +0.22 Qb4-f4 Rd8-f8 Qf4-d6 Rf8-e8 Qd6-c7 Qh8-b2 Qc7xa7 Qb2-c1 Rd5-e5 Nc5-e6 Qa7xb6 Qc1xh6 Qb6-b4 Kg8-h8 Qb4-d6 g6-g5 Kg2-g1 Kh8-g7 Re5-d5 g5-g4 Rf3-e3 Ne6-g5 Qd6xh6+ Kg7xh6 Re3xe7 Re8xe7 Kg1-g2 Re7-e2 Rd5-d6+ f7-f6 Rd6xf6+ Kh6-g7 Rf6-d6 Re2xf2+ Kg2xf2 Ng5-e4+ Kf2-e3 Ne4xd6 Ke3-f4 Nd6-b5 Bb3-a4 Nb5-c3 Ba4xd7 Nc3xa2 Kf4xg4 Na2-b4 Bd7-f5

So of course SF does not see it as a best move. But who is to say what move is without analyzing about 29 possible candidates. But I guess this is where a couple of thousand cores come in handy because if you can figure out a way to simultaneously search a bunch of candidate moves and then at the same time go another search further with each of the candidates best moves, then you might be putting those 2000 cores to a great use and you would find moves like Rxc5 pretty easily.

But was it really a winning move as great as it looks with a 10 out of 10 WOW factor? Who knows....

If you play the game forward to this position after 1. Rxc5 bxc5 2. Qh4 Rde8 3. Rf6

[fen]4r1kq/p2prp1p/5RpP/2p5/7Q/1B4P1/P4PK1/8 w - - 3 3[/fen]

In this position SF 8 played 3. ... Rf8?

Well all I can say is that again there seems to be a cache problem or something as SF8 on my computer shows this as the

26/56 00:05 48,715k 9,207k -8.17 Re8-f8 Qh4-f4 d7-d5 Bb3xd5 Re7-d7 Bd5-c4 Rd7-c7 Qf4-d6 Rf8-e8 Bc4-d5 c5-c4 g3-g4 c4-c3 g4-g5 Qh8xf6 Qd6xf6 Kg8-f8 Qf6-g7+ Kf8-e7 Bd5xf7 Ke7-d8 Bf7xe8 Rc7xg7 h6xg7 c3-c2 Be8xg6 c2-c1Q g7-g8Q+ Kd8-c7 Qg8-g7+ Kc7-d8 Qg7-h8+ Kd8-d7 Qh8-e8+ Kd7-d6 Qe8-d8+ Kd6-e6 Qd8-f6+ Ke6-d7 Qf6-f5+ Kd7-d8 Qf5-f8+ Kd8-c7 Qf8-e7+ Kc7-b8 Qe7-e5+ Qc1-c7 Qe5xc7+ Kb8xc7
26/56 00:05 48,715k 9,207k -7.76 Re8-a8 Qh4-f4 Re7-e6 Rf6xf7 Qh8-d4 Rf7-g7+ Qd4xg7 h6xg7 Kg8xg7 Bb3xe6 d7xe6 Qf4-e5+ Kg7-f7 Qe5xc5 Kf7-f6 Qc5-d4+ Kf6-e7 Qd4-h4+ Ke7-d7 Qh4xh7+ Kd7-d6 Qh7xg6 Ra8-c8 Qg6-d3+ Kd6-e7 Qd3-h7+ Ke7-d6 Qh7xa7 e6-e5 Qa7-b6+ Rc8-c6 Qb6-b4+ Kd6-d5 Qb4-b5+ Kd5-d6 Qb5-d3+ Kd6-e6 Qd3-e4 Rc6-c3 a2-a4 Ke6-f6 a4-a5
26/56 00:05 48,715k 9,207k -7.51 a7-a5 Qh4-f4 Re7-e6 Rf6xf7 Qh8-d4 Rf7-g7+ Qd4xg7 Bb3xe6+ Re8xe6 h6xg7 Kg8xg7 Qf4-c7 a5-a4 Qc7xd7+ Kg7-f6 Qd7xh7 Kf6-e5 Qh7-c7+ Ke5-d5 Qc7-d7+ Re6-d6 Qd7xa4 Rd6-b6 Qa4-a8+ Kd5-c4 Qa8-a5 Rb6-b4 Qa5-a6+ Kc4-d5 Qa6xg6 c5-c4 Qg6-g5+ Kd5-c6 Qg5-a5 Rb4-b5 Qa5-c3 Rb5-c5 f2-f4 Kc6-b5
26/56 00:05 48,715k 9,207k -7.41 Re8-c8 Qh4-f4 d7-d5 Bb3xd5 Kg8-f8 Bd5xf7 Qh8xf6 Qf4xf6 Re7xf7 Qf6-h8+ Kf8-e7 Qh8xc8 Ke7-d6 Qc8-a6+ Kd6-d5 Qa6-d3+ Kd5-c6 f2-f4 Rf7-c7 Kg2-f3 Kc6-b6 Qd3-c4 a7-a6 g3-g4 a6-a5 f4-f5 g6xf5 g4xf5 Rc7-c6 Qc4-b3+ Kb6-a6 a2-a4 Rc6xh6 Qb3-b5+ Ka6-a7 Qb5xc5+ Ka7-a6 Qc5-b5+ Ka6-a7 Qb5xa5+ Ka7-b7 Qa5-e5
26/56 00:05 48,715k 9,207k -6.57 Re8-b8 Qh4-f4 Rb8xb3 a2xb3 Kg8-f8 Rf6xf7+ Re7xf7 Qf4-b8+ Kf8-e7 Qb8xh8 d7-d6 Qh8-c8 Ke7-f6 f2-f4 Rf7-e7 Qc8-c6 Kf6-e6 Kg2-f3 Re7-f7 Kf3-e4 Rf7-e7 Qc6-c8+ Re7-d7 Qc8-e8+ Rd7-e7 Qe8-c6 a7-a6 Ke4-f3 Re7-f7 Qc6-e4+ Ke6-d7 Qe4-b7+ Kd7-e6 Qb7-c6 Rf7-e7 Kf3-e4 a6-a5 Qc6-c8+ Re7-d7 Qc8-e8+ Rd7-e7 Qe8-c6 Re7-d7 Qc6-d5+ Ke6-f6 Qd5-g5+ Kf6-f7 Ke4-d5
26/56 00:05 48,715k 9,207k -5.97 Re7-e6 Bb3xe6 d7xe6 g3-g4 Kg8-f8 g4-g5 Re8-d8 Qh4-f4 Kf8-e8 Rf6xf7 Qh8-d4 Qf4xd4 Rd8xd4 Rf7xh7 c5-c4 Rh7-c7 Rd4-g4+ Kg2-f3 Rg4xg5 h6-h7 Rg5-h5 h7-h8R+ Rh5xh8 Rc7-c8+ Ke8-d7 Rc8xh8 Kd7-d6 Kf3-e4 Kd6-c5 Rh8-c8+ Kc5-b4 Ke4-d4 c4-c3 Rc8-c4+ Kb4-a3 Kd4xc3 e6-e5 Rc4-c5 Ka3xa2 Rc5xe5
26/56 00:05 48,715k 9,207k -3.89 c5-c4 Bb3xc4 g6-g5 Qh4xg5+ Kg8-f8 Bc4-d5 Qh8-g8 Qg5-f4 Re7-e6 Rf6xe6 d7xe6 Bd5-c6 Qg8-g6 Qf4-b4+ Kf8-g8 Bc6xe8 Qg6xh6 Be8-c6 Qh6-c1 Bc6-f3 Qc1-c7 Qb4-g4+ Kg8-f8 Qg4-h5 Kf8-g7 Qh5-g5+ Kg7-h8 a2-a4 Qc7-d6 Bf3-e4 Qd6-f8 Qg5-h4 f7-f5 Be4-d3 a7-a5 Qh4-d4+ Kh8-g8 Qd4-e5 Qf8-c8 Qe5xa5 Qc8-b7+ Kg2-g1
26/56 00:05 48,715k 9,207k -3.41 d7-d5 Bb3xd5 Kg8-f8 Qh4-f4 Qh8-g8 Qf4-d6 g6-g5 Qd6xc5 Re8-b8 g3-g4 Rb8-d8 Bd5-b3 a7-a5 Bb3xf7 Qg8xf7 Qc5xg5 a5-a4 Qg5-g7+ Kf8-e8 Rf6xf7 Re7xf7 Qg7-g8+ Ke8-e7 Qg8-g5+ Ke7-d7 Qg5-d5+ Kd7-e7 Qd5-e4+ Ke7-d6 Qe4-d4+ Kd6-e7 Qd4xa4 Rd8-f8 Qa4-e4+ Ke7-d8 Kg2-g3 Rf8-e8 Qe4-d5+ Rf7-d7 Qd5-a5+ Kd8-e7 g4-g5 Ke7-d6 Qa5-b6+ Kd6-d5 f2-f4 Rd7-e7
26/56 00:05 48,715k 9,207k -2.41 g6-g5 Qh4xg5+ Kg8-f8 Qg5-f5 Qh8-g8 Qf5xc5 Re8-d8 Bb3xf7 Qg8xf7 Rf6xf7+ Kf8xf7 Qc5xa7 Rd8-c8 a2-a4 Rc8-c6 a4-a5 Rc6xh6 Qa7-b7 Re7-e6 Qb7xd7+ Kf7-f8 Qd7-d8+ Kf8-f7 Qd8-c7+ Kf7-g8 Qc7-c8+ Kg8-f7 Qc8-b7+ Kf7-g8 Qb7-d5 Kg8-f8 f2-f4 Re6-a6 Qd5-d4 Kf8-f7 Qd4-d7+ Kf7-f8 Qd7-c8+ Kf8-f7 Qc8-b7+ Kf7-f8 Qb7-b8+ Kf8-f7 Qb8-b3+ Kf7-e7 Qb3-g8 Ke7-d7 Qg8-g7+ Kd7-c8 Qg7-f8+ Kc8-b7
26/56 00:05 48,715k 9,207k 0.00 Kg8-f8 Qh4-f4 Qh8-g8 Rf6xf7+ Re7xf7 Qf4-d6+ Re8-e7 Qd6-b8+ Re7-e8 Qb8-d6+

10th best move after 5 seconds and 26 ply depth. It shows that the correct move to play is 3. ... Kf8 and that everything else is bad.

[fen]4rk1q/p2prp1p/5RpP/2p5/7Q/1B4P1/P4PK1/8 w - - 4 4[/fen]

I don't think that you can convince me that Black doesn't have a chance to draw the game after 3. ... Kf8. In fact 3. ... Rf8 shows up as -8.17

It also shows up as -8.18 after 3 seconds and 25 ply.

25/56 00:03 30,456k 9,339k -8.18 Re8-f8 Qh4-f4 d7-d5 Bb3xd5 Re7-d7 Bd5-c4 Rd7-c7 Qf4-d6 Rf8-e8 Bc4-d5 c5-c4 g3-g4 c4-c3 g4-g5 Qh8xf6 Qd6xf6 Kg8-f8 Qf6-g7+ Kf8-e7 Bd5xf7 Ke7-d8 Bf7xe8 Rc7xg7 h6xg7 c3-c2 Be8xg6 c2-c1Q g7-g8Q+ Kd8-c7 Qg8-g7+ Kc7-d8 Qg7-h8+ Kd8-d7 Qh8-e8+ Kd7-d6 Qe8-d8+ Kd6-e6 Qd8-f6+ Ke6-d7 Bg6-f5+ Kd7-e8 Qf6-e5+ Ke8-f8 Qe5-h8+ Kf8-e7 Qh8xh7+ Ke7-d6

Well this game is a win as a result of SF 8 bugs and not because Alpha forced a win with the move Rxc5. The move itself was interesting though.

ps... if you count 5 ply + 25 ply = 30 ply then you will quickly see that this should all have been analyzed by SF8 to begin with as 30 play takes only about a minute.

Buggy SF 8 regards .........
Nick
herO
Member
Posts: 342
Joined: Wed Nov 12, 2014 8:49 am

Post by herO »

AlphaZero won 28 games against Stockfish, not only 8 games.
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

herO wrote:AlphaZero won 28 games against Stockfish, not only 8 games.
It doesn't matter if it is a thousand. What has come to light for me is that SF8 does not analyze all the move options correctly.

Since pretty much every other program is based around the same codes you have to assume that bugs are everywhere.

You can give Alpha credit though for bringing this to light.

Lots of clean up work coming next for all the world's chess programmers :) They should have learned from Sargon 3 instead of Fruit .. LMAO.

Best regards
Nick
User avatar
Fernando
Admiral of the Fleet
Posts: 3059
Joined: Tue Jul 31, 2007 4:35 pm
Location: Santiago de Chile

Post by Fernando »

I my opinion what is amazing about this Google monster is the fact it learned the game and got such a level in just a couple of hours.
Any engine is at the end the result of centuries of chess playing, slowly accumulated knowledge, etc, which in one way or another get embedded in the engines that add to that, of course, his calculation power.
I cannot care less if this monster is better or not than any existing engine, but the fact of how fast it got what he can do. I wonder what will happens when mosnter ike this but even more powerfull be given not just the dabatase and rules of a game, but lot more, in extremis, the full knowledge in any field accumulated by the human race. I think we can debate about this...
Festina Lente
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

Fernando wrote:I my opinion what is amazing about this Google monster is the fact it learned the game and got such a level in just a couple of hours.
Any engine is at the end the result of centuries of chess playing, slowly accumulated knowledge, etc, which in one way or another get embedded in the engines that add to that, of course, his calculation power.
I cannot care less if this monster is better or not than any existing engine, but the fact of how fast it got what he can do. I wonder what will happens when mosnter ike this but even more powerfull be given not just the dabatase and rules of a game, but lot more, in extremis, the full knowledge in any field accumulated by the human race. I think we can debate about this...
That's the problem though no one has those 2000 core computers. Maybe you have to wait 5-10 years when they become available.

Just like IBM in the past this could be a Marketing ploy to push new generations of computers our way.

OH btw... by then all the amateur programmers will have figured out the same and you will have loads to choose from again for analysis. Since playing them will remain a waste of time unless you like the pain of losing for ever.

Best regards
Last edited by spacious_mind on Sun Dec 17, 2017 11:54 pm, edited 1 time in total.
Nick
User avatar
Fernando
Admiral of the Fleet
Posts: 3059
Joined: Tue Jul 31, 2007 4:35 pm
Location: Santiago de Chile

Post by Fernando »

spacious_mind wrote:
Fernando wrote:I my opinion what is amazing about this Google monster is the fact it learned the game and got such a level in just a couple of hours.
Any engine is at the end the result of centuries of chess playing, slowly accumulated knowledge, etc, which in one way or another get embedded in the engines that add to that, of course, his calculation power.
I cannot care less if this monster is better or not than any existing engine, but the fact of how fast it got what he can do. I wonder what will happens when mosnter ike this but even more powerfull be given not just the dabatase and rules of a game, but lot more, in extremis, the full knowledge in any field accumulated by the human race. I think we can debate about this...
That's the problem though no one has those 2000 core computers. Maybe you have to wait 5-10 years when they become available.

Just like IBM in the past this could be a Marketing ploy to push new generations of computers our way.

OH btw... by then all the amateur programmers will have figure out the same and you will have loads to choose from again for analysis. Since playing them will remain a waste of time unless you like the pain of losing for ever.

Best regards
BTW, this week I will mail you this weird program called ALEX by a gentleman named Van Tiggelen. It is....well, you will see. MAYBE you can get something from it.
Festina Lente
User avatar
spacious_mind
Senior Member
Posts: 3999
Joined: Wed Aug 01, 2007 10:20 pm
Location: Alabama
Contact:

Post by spacious_mind »

Fernando wrote:
spacious_mind wrote:
Fernando wrote:I my opinion what is amazing about this Google monster is the fact it learned the game and got such a level in just a couple of hours.
Any engine is at the end the result of centuries of chess playing, slowly accumulated knowledge, etc, which in one way or another get embedded in the engines that add to that, of course, his calculation power.
I cannot care less if this monster is better or not than any existing engine, but the fact of how fast it got what he can do. I wonder what will happens when mosnter ike this but even more powerfull be given not just the dabatase and rules of a game, but lot more, in extremis, the full knowledge in any field accumulated by the human race. I think we can debate about this...
That's the problem though no one has those 2000 core computers. Maybe you have to wait 5-10 years when they become available.

Just like IBM in the past this could be a Marketing ploy to push new generations of computers our way.

OH btw... by then all the amateur programmers will have figure out the same and you will have loads to choose from again for analysis. Since playing them will remain a waste of time unless you like the pain of losing for ever.

Best regards
BTW, this week I will mail you this weird program called ALEX by a gentleman named Van Tiggelen. It is....well, you will see. MAYBE you can get something from it.
Send me your address maybe when I have time in the New Year I will prepare a laptop and send it to you so you can play. Last time I went to a used computer store I bought a couple of I5 business laptops for almost no money. So if I find another then will fix it up with emulators that work :)

Best regards
Nick
User avatar
Fernando
Admiral of the Fleet
Posts: 3059
Joined: Tue Jul 31, 2007 4:35 pm
Location: Santiago de Chile

Post by Fernando »

Dear Nick, no neccesary all that work
I am happy with what I have.
We are not making Quid pro quo business. I send this program as the previous as a gift and happy to do so, not as a reimburse of a past or future delivery.
The sargon 3 failed probably because of me and so it is OK, You did an enormous effort and I am very grateful for that.
Festina Lente
Post Reply