The Doomsday of Stockfish DD

[Lyudmil Tsvetkov]

You are playinng definitely better than 2000 ELO. If I have seen just your games I would bet something between 2250 - 2350 FIDE. Anyway, playing against human opponents is something completly different and I didn´t seen your human vs human games

There is a small sample of my human games here: http://www.365chess.com/players/Ludmil_Tsvetkov , dating back from the year 2004. I really played very few competitive games, mostly for Bulgarian rating, even fewer for FIDE, as I have been professionally occupied. The engine mania however never left me.

[overlord]

Two hours ago I have installed Stockfish DD... here is the first (little bit boring draw).

[Event "Blitz (5 minutes per game)"]
[Site "?"]
[Date "2013.12.05"]
[Round "?"]
[White "Kvicala, Miroslav"]
[Black "Stockfish DD"]
[Result "1/2-1/2"]
[PlyCount "93"]

1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. O-O Nxe4 5. Re1 Nd6 6. Nxe5 Be7 7. Nxc6 dxc6 8. Bf1 O-O 9. d4 Re8 10. c3 Bf6 11. Bf4 Rxe1 12. Qxe1 Bg5 13. Qe3 Bxf4 14. Qxf4 Be6 15. Nd2 Qd7 16. a3 Re8 17. Re1 f6 18. g3 Bf7 19. Rxe8+ Qxe8 20. Qe3 Qe6 21. Qxe6 Bxe6 22. f4 Kf7 23. Bd3 g6 24. h4 Bf5 25. Bxf5 Nxf5 26. Kf2 h5 27. Ne4 Nd6 28. Nxd6+ cxd6 29. Ke3 b5 30. b3 Ke6 31. Ke4 a5 32. a4 bxa4 33. bxa4 d5+ 34. Ke3 Kd6 35. Kf3 c5 36. Ke3 c4 37. Kf3 Ke6 38. Ke3 Kf5 39. Kf3 g5 40. fxg5 fxg5 41. hxg5 Kxg5 42. Kf2 Kg4 43. Kg2 Kf5 44. Kf3 Kg5 45. Kf2 Kg4 46. Kg2 Kf5 47. Kf3 1/2-1/2

[Daniel Shawul]

Yak, you exchanged all pieces This can be easily avoided by having contempt, because right now it is assuming you are a 3200 elo player. But then it would play badly against equally ranked engines. Some engines happily use contempt to bloat their elos however. Fortresses can also be solved by engines through evaluation, but again it is not worth the effort. Your chances of winning 'properly' against such an engine remain in the order of 1 in 1000, if you are a 2000+ elo ranked player.

[duncan]

Hi Carl.

..


I am sure GMs could do much better against engines, but they need a thorough preparation, involving maybe years of time. No one has simply thought of that.

milov 's thoughts on this after defeating a handicapped rybka
http://rybkaforum.net/cgi-bin/rybkaforu ... l?tid=7517
..

My first conclusions, just after the match:

As I said before, Rybka is very strong and is the strongest engine in the world. She is strong not only in the open positions as some might think but also in defence and that`s probably the reason why it`s so hard to win against her even having f7-pawn handicap.

However, I`m convinced that the program is not unbeatable (when talking about playing normal chess without handicap). In the match we played now the question basically was whether I can convert my material advantage or not and in a way it became a little boring towards the end. I personally would be interested to play a match against this program where the only handicap I`m asking for is that I get White in all games. I`m confident that with the right preparation I`ll be able to represent the human side as well as I did in this match.

[carldaman]

Now that Marco enjoys to see Stockfish losing, I will post some more games in that vein. They are all with white, please do not ask me why, I just considered my chances higher when playing the white side and somehow played too few games with black against DD.

...

The conclusion could be that Stockfish underestimates opponent initiative with sacrifices, still slightly underestimates chains close to the enemy king, as well as storming/attacking pawns on the 6th rank when blocked.

I hope Marco and his collaborators are taking notes. I'm optimistic that the eval could be improved so that the above shortcomings are corrected without hurting the engine's overall performance in all types of positions.

I can go even as far as saying that it is worth incorporating such evaluation changes regardless, even if the overall rating does not go up, as long as the engine's play becomes more well-rounded, with visibly better play in such closed positions with flank attacks facing the king. I suspect most programmers don't agree with me on this, but SF is not even commercial, so if anyone were to try what I'm proposing, it could be them (the SF team).

At the very least, a special positionally enhanced version of Stockfish could exist, as a companion to the 'normal' (strongest) version.

Nice games and attacking ideas, Lyudmil!

CL

[carldaman]

Hi Carl, thanks for the games.

That is the problem with many test suites and opening books: they contain an awful percentage of unequal positions, just like the one you used. 9. Bd3 is very bad seemingly, somewhere to the level of Nunn . For me, after black's f3 in both games, white is hopeless. Probably the only white chance to hold the game would have been playing f3 itself. Interesting that Komodo does not find it at such a long TC.

You conveyed a point; nice comments btw.

I see what you mean regarding the opening, but while this position is not the best line White could have played (9.Bd3 is probably ?!), I can say that
a) White is not (much) worse objectively and should be able to draw with correct play for both sides
b)we are able to arrive at such an assessment only after much testing and analysis; John Nunn certainly must've liked his chances when he embarked upon this line with White

As a tester, I like this variation a lot since it reveals so much about the 'character' of each engine, and both strengths and weaknesses immediately can bubble to the surface after a smallish number of games.

For the sake of comparison I shall include the corresponding games SF played with White, showing how it was able to evade the pitfalls:

[pgn]

[Event "15m3s-SFsbK6 test"]
[Site "?"]
[Date "2013.11.29"]
[Round "2.1"]
[White "Stockfish 241113 64 SSE4.2sb"]
[Black "Komodo 6 64-bit"]
[Result "1/2-1/2"]
[ECO "B32"]
[Annotator "0.16;0.14"]
[PlyCount "70"]
[EventDate "2013.11.29"]
[EventType "tourn"]

{Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz 3292 MHz W=26.4 plies; 1,539kN/s
B=21.3 plies; 1,077kN/s} 1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e5 5. Nb5 d6 6.
c4 Be7 7. N1c3 a6 8. Na3 f5 9. Bd3 f4 10. O-O {[%eval 16,24] [%emt 0:00:29]}
Nf6 {[%eval 14,21] [%emt 0:00:23]} 11. Nd5 {[%eval 10,24] [%emt 0:00:24]} O-O {
[%eval 11,20] [%emt 0:00:18]} 12. Nc2 {[%eval 0,25] [%emt 0:00:16] (Bd2)} Nxd5
{[%eval 0,22] [%emt 0:00:35]} (12... f3 $2 13. Nxe7+) 13. cxd5 {[%eval 0,26]
[%emt 0:00:21] (exd5)} Nb8 {[%eval 0,22] [%emt 0:00:22]} 14. b4 {[%eval 0,26]
[%emt 0:00:13] (Bd2)} (14. Ne1) (14. a4) (14. b3 $5) 14... Qe8 {[%eval 0,21]
[%emt 0:00:35] (Nd7)} (14... f3 $5 $13) 15. Kh1 {[%eval 0,23] [%emt 0:00:27]
(Ne1)} Nd7 {[%eval -17,19] [%emt 0:00:37]} 16. Bd2 {[%eval 0,25] [%emt 0:00:26]
(a4)} Rf6 {[%eval -24,19] [%emt 0:00:40] (Nb6)} 17. Rg1 {[%eval 0,25] [%emt 0:
00:22] (Rc1)} a5 {[%eval -22,19] [%emt 0:00:15] (Rf8)} 18. a3 {[%eval 22,23]
[%emt 0:00:15]} Bd8 $5 {[%eval -22,20] [%emt 0:00:32] (axb4)} (18... Qf7 $5 19.
Qe2 Rh6 $15) (18... Rh6) (18... axb4) 19. Qe2 {[%eval -6,22] [%emt 0:00:25]}
Bb6 {[%eval -30,20] [%emt 0:00:30]} 20. Bb5 {[%eval 0,24] [%emt 0:00:16] (Rac1)
} Qf8 $6 {[%eval -25,20] [%emt 0:00:41] (Qf7)} (20... Qe7 $5 {planning} 21. --
Rh6 $40 22. -- Rxh2+ $3) (20... Qd8 $5 21. bxa5 Rh6 22. g4 $5 (22. g3 Bxa5 23.
f3 Bxd2 24. Qxd2 Nc5 25. Rab1 fxg3 26. Rxg3 Qh4 $15) 22... Bxa5 $15 (22... Rh3
$5 23. axb6 f3 24. Qf1 Qh4 25. Qxh3 Qxh3 26. Bxd7 Bxd7 27. Bg5 $1 Bxg4 28. Ne3
$44) 23. Bb4 Bb6 (23... Qh4 24. f3 Bb6 25. Rg2 Nc5 26. a4 $13) 24. g5 $13 Rh4
25. Bxd6 Nf6 $1 26. Bxe5 Ng4 27. Rxg4 Bxg4 28. f3 Qxg5 (28... Bh3 $13 29. Bd4
Bxd4 30. Nxd4 Qxg5 31. Qb2 Rf8 32. Rg1 Qf6 (32... Qh5 33. Rxg7+) 33. Rb1 $44)
29. Bd4 $8 Bxd4 30. Nxd4 Qe5 $1 31. Rd1 $8 Bh3 $15) 21. bxa5 {[%eval -22,25]
[%emt 0:01:21] (Ne1)} Bxa5 {[%eval -35,19] [%emt 0:00:12]} 22. Bxa5 {[%eval
-24,25] [%emt 0:00:13] (Bb4)} Rxa5 {[%eval -48,17] [%emt 0:00:07]} 23. Bxd7 {
[%eval -42,24] [%emt 0:00:21] (a4)} Bxd7 {[%eval -42,17] [%emt 0:00:06]} 24. f3
{[%eval -42,27] [%emt 0:00:25] (Qd2)} Rc5 {[%eval -35,21] [%emt 0:00:52] (Qa8)}
25. Nb4 {[%eval -22,26] [%emt 0:00:30] (Rab1)} Bb5 {[%eval -28,22] [%emt 0:01:
22] (Qd8)} 26. Qd2 {[%eval -16,26] [%emt 0:00:24]} Rf7 {[%eval -14,23] [%emt 0:
01:15] (Qc8)} 27. a4 {[%eval -2,26] [%emt 0:00:14] (Nd3)} Bc4 {[%eval -4,22]
[%emt 0:00:22]} 28. Rgc1 {[%eval 0,28] [%emt 0:00:13] (Nd3)} Qa8 {[%eval 0,21]
[%emt 0:00:18] (Qd8)} 29. Nd3 {[%eval 2,27] [%emt 0:00:19]} Rcc7 {[%eval 0,21]
[%emt 0:00:09]} 30. h3 {[%eval 10,26] [%emt 0:00:18]} Qc8 {[%eval 0,22] [%emt
0:00:20] (Qd8)} 31. Kh2 {[%eval 0,28] [%emt 0:00:11] (Rab1)} h6 {[%eval 0,23]
[%emt 0:00:20]} 32. Rab1 {[%eval 0,31] [%emt 0:00:15] (Rc3)} Kh7 {[%eval 0,24]
[%emt 0:00:19] (Bxd3)} 33. Rc3 {[%eval 0,30] [%emt 0:00:13]} Bxd3 {[%eval 0,25]
[%emt 0:00:10]} 34. Rxc7 {[%eval 0,34] [%emt 0:00:11]} Rxc7 {[%eval 0,27]
[%emt 0:00:07]} 35. Qxd3 {[%eval 0,37] [%emt 0:00:12]} Rc2 {[%eval 0,28] [%emt
0:00:09]} (35... Rc2 36. Qb3 Qc5 37. Qb6 $1 $11 (37. Qxb7 $2 Qf2 $1 $19 38. Rg1
Qxf3)) 1/2-1/2

[Event "25m5s-SFsbK6 test"]
[Site "?"]
[Date "2013.11.29"]
[Round "2.1"]
[White "Stockfish 241113 64 SSE4.2sb"]
[Black "Komodo 6 64-bit MP"]
[Result "1/2-1/2"]
[ECO "B32"]
[Annotator "0.08;0.07"]
[PlyCount "88"]
[EventDate "2013.11.29"]
[EventType "tourn"]

{Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz 3292 MHz W=29.0 plies; 2,731kN/s; 3,
810 TBAs B=23.7 plies; 2,053kN/s} 1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e5 5.
Nb5 d6 6. c4 Be7 7. N1c3 a6 8. Na3 f5 9. Bd3 f4 10. O-O {[%eval 8,26] [%emt 0:
00:31]} Nf6 {[%eval 7,21] [%emt 0:00:15]} 11. Nd5 {[%eval 16,26] [%emt 0:00:27]
} O-O {[%eval 9,22] [%emt 0:00:28]} 12. Bd2 {[%eval 0,25] [%emt 0:00:23]} Be6 {
[%eval 4,20] [%emt 0:00:27]} (12... Qe8 {!!? of course could be tried here} 13.
Nc7 Qg6 $40 {best move would be} 14. Bxf4 $1 (14. Nxa8 $140 $2 Bh3 $19) 14...
Bg4 15. f3 Rac8 $13 {with a complicated and unclear position, but at least
equal for black} 16. fxg4 Nxg4 17. Bg3 (17. Nd5 $143 exf4 18. Nc2 $143 f3 $1
19. Rxf3 Bh4 $17 {threatens} 20. -- Nf2) 17... Rxf1+ 18. Qxf1 Rxc7 $13) 13. Nc2
{[%eval 2,25] [%emt 0:00:43] (Bc3)} Rc8 {[%eval 0,21] [%emt 0:00:58]} 14. Rc1 {
[%eval 8,26] [%emt 0:01:03] (Bc3)} Qe8 {[%eval 5,20] [%emt 0:00:48]} 15. Nxe7+
{[%eval 16,26] [%emt 0:00:56] (f3)} Qxe7 {[%eval 0,21] [%emt 0:00:20]} 16. f3 {
[%eval 12,27] [%emt 0:00:30] (Nb4)} Nd7 {[%eval 0,21] [%emt 0:00:30] (Qd8)} (
16... g5 $5) 17. Be1 {[%eval 30,24] [%emt 0:00:31]} h5 {[%eval 5,22] [%emt 0:
00:57] (Nc5)} 18. Be2 {[%eval 34,26] [%emt 0:01:25] (Nb4)} h4 {[%eval 10,22]
[%emt 0:00:32]} 19. Kh1 {[%eval 51,26] [%emt 0:00:34]} h3 {[%eval 23,23] [%emt
0:01:51] (g5) Komodo often likes to push its h-pawn up, but this plan seems
ineffective after having castled K-side} 20. gxh3 {[%eval 44,27] [%emt 0:00:23]
} Bxh3 {[%eval 12,23] [%emt 0:00:29]} 21. Rg1 {[%eval 36,27] [%emt 0:00:24]}
Qe8 {[%eval 11,23] [%emt 0:00:46] (Kf7)} 22. Nb4 {[%eval 36,26] [%emt 0:01:08]
(Bb4)} ({SF refrains from} 22. Qxd6 Rf6 23. Qd5+ Kh7 24. Bc3 Qh5 (24... Nb6 $5
25. Qd2 Na4 26. b4 (26. Qe1 Rh6 27. Nb4 Nxc3 28. bxc3 Qh5 29. Qf2 Be6 $44)
26... Nxc3 27. Qxc3 Qh5 28. Qe1 Ne7 $5 (28... Rh6 29. Qf2 Rh8 30. b5) 29. c5
Rh6 30. Rd1 Be6 31. Rg2 Bh3 32. Rf2 a5 33. Bf1 $1 (33. a3 $143 b6 $1 34. Ba6
Rc7 35. cxb6 $4 Rxc2 $1 36. Rxc2 Qxf3+) 33... axb4 34. Bxh3 Qxh3 35. Nxb4 Nc6
36. Nd3 Qh5 $44) 25. Qd2 Rh6 26. Bd1 a5 $5 $13 27. Qf2 b6 28. Ne1 Nc5 29. Rc2 {
and White is holding on to extra pawn in a cramped position}) 22... Rf6 {
[%eval 0,22] [%emt 0:00:21]} 23. Bc3 {[%eval 38,27] [%emt 0:01:29] (Nd5)} Nxb4
{[%eval 6,22] [%emt 0:00:54]} 24. Bxb4 {[%eval 24,27] [%emt 0:00:42]} Nc5 {
[%eval 2,22] [%emt 0:00:28]} 25. Bf1 {[%eval 32,29] [%emt 0:00:45]} Bxf1 {
[%eval 11,21] [%emt 0:00:22]} 26. Qxf1 {[%eval 16,28] [%emt 0:00:44]} Rh6 {
[%eval 0,21] [%emt 0:00:19] (Rg6)} 27. Rc2 {[%eval 40,27] [%emt 0:00:46] (b3)}
b6 {[%eval 0,21] [%emt 0:00:37] (Rc7)} 28. Rcg2 {[%eval 48,28] [%emt 0:00:39]}
Rc7 {[%eval 0,21] [%emt 0:00:10]} 29. Bxc5 {[%eval 32,28] [%emt 0:02:10]} bxc5
{[%eval 0,23] [%emt 0:00:18] (dxc5)} 30. Qe1 {[%eval 20,30] [%emt 0:00:18]
(Qd3)} Qh5 {[%eval 0,26] [%emt 0:00:49]} 31. Rf2 {[%eval 20,31] [%emt 0:00:25]}
Ra7 {[%eval 0,28] [%emt 0:00:55] (Qf7)} 32. Qa5 {[%eval 10,30] [%emt 0:00:22]}
Qh4 {[%eval 0,28] [%emt 0:00:31]} 33. Rgg2 {[%eval 10,32] [%emt 0:00:32] (Rfg2)
obviously, White has enough to cover h2} Qe7 {[%eval 0,27] [%emt 0:00:29]} 34.
Rg4 {[%eval 6,32] [%emt 0:00:33] (Rd2)} Qc7 {[%eval 0,28] [%emt 0:00:38]} 35.
Qa4 {[%eval 6,32] [%emt 0:00:20]} Qd7 {[%eval 0,28] [%emt 0:00:42] (a5)} 36.
Qd1 {[%eval 6,31] [%emt 0:00:16] (Qa5)} a5 {[%eval 0,27] [%emt 0:00:35]} 37.
Rfg2 {[%eval 6,31] [%emt 0:00:17]} Kh7 {[%eval 0,27] [%emt 0:00:27] (Qe6)} 38.
b3 {[%eval 1,32] [%emt 0:00:49] (Qb3)} Qe7 {[%eval 0,26] [%emt 0:00:23] (Qe6)}
39. Rg1 {[%eval 1,32] [%emt 0:00:24]} Qe8 {[%eval 0,26] [%emt 0:00:28] (Qc7)}
40. a3 {[%eval 0,30] [%emt 0:00:17] (R1g2)} Qc8 {[%eval 0,23] [%emt 0:00:29]
(Qe6)} 41. a4 {[%eval 0,34] [%emt 0:00:15] (Qc2)} Rb7 {[%eval 0,28] [%emt 0:00:
23] (Qd7)} 42. R1g2 {[%eval 0,36] [%emt 0:00:18] (Qd2)} Qc7 {[%eval 0,26]
[%emt 0:00:25] (Ra7)} 43. Rg5 {[%eval 0,35] [%emt 0:00:19] (Rg1)} Qb6 {[%eval
0,26] [%emt 0:00:32] (Qf7)} 44. Qg1 {[%eval 0,35] [%emt 0:00:13] (Qd3)} g6 {
[%eval 0,25] [%emt 0:00:14] (Qc7)} (44... g6 45. Rxg6 $2 Rxg6 46. Rxg6 Qxb3 $19
47. Qg4 (47. Rg8 Qxf3+) 47... Qd1+ 48. Kg2 Rb2+ 49. Kh3 Qf1+ 50. Kh4 Rxh2+)
1/2-1/2

[/pgn]

Cheers,
CL

[mcostalba]

The conclusion could be that Stockfish underestimates opponent initiative with sacrifices, still slightly underestimates chains close to the enemy king, as well as storming/attacking pawns on the 6th rank when blocked.

Sorry again for posting only white games, but that is how things happen.

Thanks for posting ! Yes they are nice to see. I guess that for better evaluation on such position perhaps a bit more of search depth (i.e. time) could be beneficial. Perhaps if you move to 5m+2s time control you can see a better Stockfish.

[zamar]

The line you are playing more or less "King's Indian" with reversed colors. It's a well known issue that most (if not all) engines still misplay KID types of positions. Mating threats are just too deep behind the horizon.

[Lyudmil Tsvetkov]

The line you are playing more or less "King's Indian" with reversed colors. It's a well known issue that most (if not all) engines still misplay KID types of positions. Mating threats are just too deep behind the horizon.

That is what the evaluation is for: to rely on it when you see nothing.

[Lyudmil Tsvetkov]

Thanks for posting ! Yes they are nice to see. I guess that for better evaluation on such position perhaps a bit more of search depth (i.e. time) could be beneficial. Perhaps if you move to 5m+2s time control you can see a better Stockfish.

Hi Marco.

Please, do not get upset. In due time I will post also some wonderful Stockfish games. This message is a one-side affair, but everyone knows how strong Stockfish is, let us see a bit of the other side.

I am sure at 5 + 2 Stockfish will do better, but, unfortunately, if I allow Stockfish that much time, I will have to play 3 times fewer games...I prefer to play more games. Maybe sometime I will play a longer time control match, but not now.