Just discovered a Rybka 3 bug

[Uri Blass]

1)wrong bishop is a case that you have a position like the following

[d]7k/1p6/1p6/8/7P/8/2B4P/4K3 w - - 0 1
Rybka evaluates it wrong.

2)considering the game that you give not only toga and shredder cannot reproduce the move but also rybka.

No 89.Qe1 by rybka3 and no 90.Qh1 by Rybka3.

Phil Harris2,R - Jefk,R, 5m + 0s, rated blitz 2009
8/5r2/5bk1/6r1/1Bp5/P4PQ1/KP2p3/8 w - - 0 1

Analysis by Rybka 3 1-cpu 32-bit :

89.Qg3-f2
= (0.00) Depth: 2 00:00:00
89.Qg3-f2
= (0.00) Depth: 3 00:00:00
89.Qg3-f2
= (0.00) Depth: 4 00:00:00
89.Qg3-f2 Rg5-e5
= (0.00) Depth: 5 00:00:00
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 6 00:00:00 2kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 7 00:00:00 3kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 8 00:00:00 9kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 9 00:00:00 15kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 10 00:00:00 26kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 11 00:00:01 59kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 12 00:00:03 117kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 13 00:00:08 315kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 14 00:00:12 469kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 15 00:00:20 790kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 16 00:00:39 1552kN
89.Qg3-f2 Rg5-e5 90.Qf2-g3+
= (0.00) Depth: 17 00:01:05 2580kN

(so k, 02.05.2009)

Phil Harris2,R - Jefk,R, 5m + 0s, rated blitz 2009
8/5r2/5bk1/8/1Bp5/P4P2/KP2p1r1/4Q3 w - - 0 1

Analysis by Rybka 3 1-cpu 32-bit :

90.Qe1-h1
= (0.00) Depth: 2 00:00:00
90.Qe1-h1
= (0.20) Depth: 3 00:00:00
90.Qe1-h1
² (0.40) Depth: 3 00:00:00
90.Qe1-h1
² (0.46) Depth: 3 00:00:00
90.Qe1-b1+
= (0.00) Depth: 4 00:00:00
90.Qe1-b1+ Kg6-g7
= (0.00) Depth: 5 00:00:00
90.a3-a4 Bf6-e5 91.Qe1-b1+
³ (-0.42) Depth: 6 00:00:00 17kN
90.a3-a4 Bf6-e5 91.Qe1-b1+ Kg6-g7
³ (-0.41) Depth: 7 00:00:00 22kN
90.a3-a4 Rf7-b7 91.Bb4-a3 c4-c3 92.Qe1-b1+
µ (-0.91) Depth: 8 00:00:02 78kN
90.Qe1-b1+ Kg6-g7
³ (-0.67) Depth: 8 00:00:02 82kN
90.Qe1-b1+ Kg6-g7 91.Qb1-e1 Rf7-d7 92.Bb4-d2 Rd7xd2 93.Qe1xd2 e2-e1Q 94.Qd2xg2+ Kg7-f8
µ (-0.85) Depth: 9 00:00:02 95kN
90.a3-a4 Rf7-d7 91.Qe1-c1 Rd7-b7 92.Qc1-c2+ Kg6-g7 93.Bb4-c3
µ (-1.01) Depth: 10 00:00:07 247kN
90.a3-a4 Rf7-d7 91.Qe1-b1+ Rd7-d3
µ (-1.25) Depth: 11 00:00:10 409kN
90.a3-a4 Rf7-d7 91.Bb4-d2 c4-c3 92.b2xc3 Rd7xd2+ 93.Qe1xd2 e2-e1Q 94.Qd2xg2+
-+ (-2.65) Depth: 12 00:00:56 2418kN
90.Qe1-c1 e2-e1Q 91.Bb4xe1 Rg2xb2+ 92.Qc1xb2 Bf6xb2 93.Ka2xb2 Rf7xf3 94.a3-a4 Kg6-f5 95.a4-a5 Kf5-e4 96.Kb2-c2 Ke4-d5 97.Be1-h4 Rf3-a3 98.Bh4-d8 Kd5-c6
-+ (-2.34) Depth: 12 00:00:59 2542kN
90.Qe1-c1 e2-e1Q 91.Bb4xe1 Rg2xb2+ 92.Qc1xb2 Bf6xb2 93.Ka2xb2 Rf7xf3 94.a3-a4 Kg6-f5 95.a4-a5 Kf5-e4 96.Kb2-c2 Ke4-d5 97.Be1-h4 Kd5-c5 98.Bh4-e1 Kc5-b5 99.Be1-d2 Rf3-f2
-+ (-2.50) Depth: 13 00:01:04 2773kN

(so k, 02.05.2009)

Uri

[M ANSARI]

This looks like a hash table bug which seems to also happen with wrong bishop. Some hash information simply is not seen until after the fact. I think this is already fixed in the new Rybka. Anyway this game was lost and no way to save it, so why not go out with a bang

Au contraire, without the bug, Rybka would've never moved 89. Qe1?, but would've drawn instead with 89. Qf2.

I am not so sure Qf2 will draw. Looks like black has too much firepower.

[M ANSARI]

This looks like a hash table bug which seems to also happen with wrong bishop.

Ansari, what do you mean by "wrong bishop"?

With wrong bishop I mean when the bishop does not control the queening square and thus is an obvious draw since opposing King can manouver in front of the passed pawn. In some positions Rybka will play a move and give an almost +6 advantage ... only to show a 0.00 immediately after the move. It is a known bug that was seen since initial release of R3 but it was not possible to fix then. I have shown this position to Vas and he seemed aware of this bug. Pretty sure it is fixed latest Rybka.

[Uri]

considering the game that you give not only toga and shredder cannot reproduce the move but also rybka.

No 89.Qe1 by rybka3 and no 90.Qh1 by Rybka3.

Well then your Rybka 3 is obviously different from my Rybka 3. Maybe you're using a different hash size than me because I'm only using 64 MB hash on an Intel Core 2 Q8200 with 4 gigs RAM.

[bnculp]

Uri Blass wrote:

considering the game that you give not only toga and shredder cannot reproduce the move but also rybka.

No 89.Qe1 by rybka3 and no 90.Qh1 by Rybka3.

Phil Harris2,R - Jefk,R, 5m + 0s, rated blitz 2009
8/5r2/5bk1/6r1/1Bp5/P4PQ1/KP2p3/8 w - - 0 1

No

I agree with Zlatnik. 89. Qe1 and 90. Qh1 are reproducible

Vista x64, Intel QX6700, 4gb ram, 1gb hash, 3-4-5-6 EGTBs

Phil Harris2,R - Jefk,R, Blitz:4'+2" Kenhorst
8/5r2/5bk1/6r1/1Bp5/P4PQ1/KP2p3/8 w - - 0 1

Analysis by Rybka 3:

89.Qf2 Re5 90.Qg3+
= (0.00) Depth: 11 00:00:00 123kN
89.Qf2 Re5 90.Qg3+
= (0.00) Depth: 12 00:00:02 329kN
89.Qe1 Re5 90.Qb1+ Kg7 91.Qg1+ Kh7 92.f4 e1Q 93.Bxe1 Re2 94.Kb1 Rxb2+ 95.Kc1 Ra2 96.Qg3
= (0.08) Depth: 12 00:00:02 329kN
89.Qe1 Re5 90.Qb1+ Kg7 91.Qg1+ Kh7 92.f4 e1Q 93.Bxe1 Re2 94.Kb1 Rxb2+ 95.Kc1 Ra2 96.Qg3
= (0.00) Depth: 13 00:00:02 389kN
89.Qe1 Re5 90.Qb1+ Kg7 91.Qg1+ Kh7 92.f4 e1Q 93.Bxe1 Re2 94.Kb1 Rxb2+ 95.Kc1 Ra2 96.Qg3 Ra1+ 97.Kc2 Ra2+ 98.Kc1 Ra1+ 99.Kc2 Ra2+ 100.Kc1 Ra1+ 101.Kc2 Ra2+ 102.Kc1 Ra1+ 103.Kc2 Ra2+ 104.Kc1
= (0.00) Depth: 14 00:00:05 816kN
89.Qe1 Re5 90.f4 Ree7 91.Qb1+ Kh6 92.Qh1+ Kg6 93.Qg1+ Kf5
= (0.14) Depth: 15 00:00:08 1426kN
89.Qe1 Re5 90.f4 Ree7 91.Qg1+ Kf5 92.Qc5+ Kxf4 93.Qxc4+ Kf3 94.Qd5+ Kg4 95.Qg2+ Kf5[] 96.Qf3+ Kg6 97.Qd3+ Kh6
= (0.20) Depth: 16 00:00:10 1864kN
89.Qe1 Re5 90.f4 Ree7 91.Qg1+ Kf5 92.Qc5+ Kxf4 93.Qxc4+ Kf3 94.Qd5+ Kg4 95.Qg2+ Kh5[] 96.Qf3+ Kg6 97.Qd3+
= (0.15) Depth: 17 00:00:23 4346kN, tb=2
89.Qe1 Re5 90.f4 Ree7 91.Qg1+ Kf5 92.Qc5+ Kxf4 93.Qxc4+ Kf3 94.Qd5+ Kg4 95.Qg2+ Kh5[] 96.Qf3+ Kg6 97.Qd3+
= (0.15) Depth: 18 00:00:32 6281kN, tb=3
89.Qe1 Re5 90.f4 Re4 91.Qg3+ Kf5 92.Qh3+ Kxf4 93.Qh6+ Kf5 94.Qh5+ Ke6[] 95.Qh3+ Kd5 96.Qf5+ Re5[] 97.Qf3+ Re4[] 98.b3 Rc7 99.Qxf6 cxb3+ 100.Kb2 Rxb4 101.Qf5+ Kd6 102.Qd3+ Ke5[] 103.Qxe2+ Re4[] 104.Qh2+
= (0.25) Depth: 19 00:02:01 23117kN, tb=65


Phil Harris2,R - Jefk,R, Blitz:4'+2" Kenhorst
8/5r2/5bk1/8/1Bp5/P4P2/KP2p1r1/4Q3 w - - 0 1

Analysis by Rybka 3:

90.Qh1
+/- (1.08) Depth: 11 00:00:00 81kN
90.Qh1
+/- (1.17) Depth: 12 00:00:01 163kN
90.Qh1
+/- (1.23) Depth: 13 00:00:02 336kN
90.Qh1
+/- (1.25) Depth: 14 00:00:04 631kN
90.Qh1
+/- (1.25) Depth: 15 00:00:05 853kN
90.Qh1
+- (1.44) Depth: 16 00:00:14 2193kN, tb=8
90.Qh1
+- (1.44) Depth: 17 00:00:23 3724kN, tb=42
90.Qh1
+- (1.44) Depth: 18 00:00:32 5044kN, tb=57
90.Qh1 Rg5 91.f4 Rg3 92.Qe4+[] Kg7 93.Qxe2 Rb3 94.Bc3 Bxc3[] 95.bxc3 Rb5 96.Qxc4 Rbf5 97.a4 Rxf4 98.Qc6 R4f6 99.Qg2+ Rg6 100.Qb2 Rf5 101.c4+ Rgf6 102.Kb3
+- (1.51) Depth: 19 00:00:54 8242kN, tb=85

[Nimzovik]

Vas has said a Rybka 3+ is being worked on............ Who knows?

[Uri Blass]

I can reproduce the problem with the multi-cpu engine but not with the single cpu engine.

The single cpu engine or the multi-cpu engine when I limit it to one processor seems to be stronger in this position.

Uri

[M ANSARI]

I can reproduce the problem with the multi-cpu engine but not with the single cpu engine.

The single cpu engine or the multi-cpu engine when I limit it to one processor seems to be stronger in this position.

Uri

Well then that would seem to put it in the category of MP hash bug. This MP hash bug definetely drops a few ELO points for Rybka 3 MP, IMHO it is worth around 20 to 30 ELO. Vas seems to think that is not so and the ELO difference is at most around 5 ELO. It sure seems like a lot more when it hits. This also explains why Rybka 3 single core is so strong in comparison to MP Rybka 3 and why the jump to MP from single core is not as high as expected.

Many were anticipating that a simple bug fix in Rybka 3 (Rybka 3+) would give a quick ELO boost. These same people are very upset that the fixed version of Rybka 3+ will end up being forced to drop its ELO rating to not better Rybka 3. Personally I look forward to Rybka 4.