Ponder hit rate of the "Litos" to Rybka 3/4

Code, algorithms, languages, construction...
User avatar
IWB
Posts: 195
Joined: Thu Jun 10, 2010 4:10 pm

Ponder hit rate of the "Litos" to Rybka 3/4

Post by IWB » Thu Jun 10, 2010 4:26 pm

Hi all

A thing which I really find very supprising and unsettling is that all the Litos have a ponder hit rate to Rybka 3 (and now R4) which is far beyond my understanding. I am not talking about 'moving fast', I am really talking about ponder hits. This is what is giving these engines an edge over Rybka. This advantage vanishes as soon as you play agains non Rybka engines. What is left then is the "normal" dominance over the opponents as Rybka is showing it.

So for those who can "read" the source code of these Litos and know a bit more about chess programing than I do: What are resonable explanations for such a behavior?

Here is one example game played without a book, ponder on(!) one thread to demonstrate my point. The game is nothing special and a draw anyhow but if you have a look how many moves are replied with 0 seconds on a 30 + 3 time control I have problems to understand that.

[Event "30 Minutes/Game + 3 Seconds/Move"]
[Site "Engine Match"]
[Date "2010.05.29"]
[Round "1"]
[White "DR 4 SSE42 x64 1T"]
[Black "Fire 1.3 POP x64 BASES 1T"]
[Result "1/2-1/2"]

1. e4 {+0.22/18 1:41m} e5 {+0.09/21 44s (e6)} 2. Nf3
{+0.16/19 2:32m} Nc6 {+0.20/22 0s} 3. d4 {+0.14/18 41s
(Bb5)} exd4 {0.00/20 42s} 4. Nxd4 {+0.14/18 0s} Nf6
{0.00/19 0s} 5. Nxc6 {+0.17/17 40s} bxc6 {+0.02/20 0s
(dxc6)} 6. Bd3 {+0.09/18 2:00m} d5 {+0.04/22 0s} 7. Nc3
{+0.11/17 1:28m} Bb4 {+0.01/22 0s} 8. exd5 {+0.12/18 54s}
cxd5 {+0.01/21 0s} 9. Qe2+ {+0.09/18 1:23m} Be7 {+0.01/22
0s} 10. O-O {+0.08/18 31s} O-O {-0.01/21 0s} 11. Bf4
{+0.08/16 30s} c6 {0.00/21 0s} 12. h3 {+0.05/16 33s} Bd6
{+0.01/21 1:17m (Re8)} 13. Qd2 {+0.01/17 1:23m (Qf3)} Rb8
{0.00/20 47s (Qc7)} 14. b3 {+0.11/19 1:36m} Re8 {0.00/22
0s} 15. Rfe1 {+0.11/17 25s} Be6 {0.00/21 0s} 16. Bxd6
{+0.12/17 29s (Rad1)} Qxd6 {-0.02/17 0s} 17. Na4 {+0.12/16
24s} c5 {-0.08/20 2:18m (Nd7)} 18. Qa5 {0.00/17 41s (Nc3)}
c4 {0.00/21 41s} 19. Bf1 {0.00/17 0s} Rec8 {0.00/19 0s}
20. Qxa7 {+0.03/18 41s} Bd7 {0.00/21 0s (Bf5)} 21. Nb2
{+0.41/15 38s} Rb6 {0.00/22 0s} 22. Rad1 {+0.36/18 42s
(Rac1)} Rc7 {0.00/23 49s} 23. Qa8+ {+0.36/18 0s} Rc8
{0.00/22 0s} 24. Qxc8+ {+0.48/19 39s (Qa7)} Bxc8 {-0.18/23
29s} 25. Nxc4 {+0.48/19 0s} dxc4 {-0.18/22 0s} 26. Rxd6
{+0.35/20 46s} Rxd6 {-0.18/24 0s} 27. Bxc4 {+0.35/20 43s}
Ba6 {-0.23/24 0s (Be6)} 28. Bxa6 {+0.53/19 1:14m} Rxa6
{-0.23/24 0s} 29. a4 {+0.53/19 15s} Kf8 {-0.23/23 5:53m}
30. f3 {+0.33/21 0s} Rc6 {-0.28/22 49s (Nd7)} 31. c4
{+0.64/19 1:02m} Rb6 {-0.06/24 0s} 32. Rb1 {+0.64/18 33s}
Ra6 {-0.06/23 0s} 33. Kf2 {+0.69/20 28s (Ra1)} Ke7 {0.00/21
24s} 34. Ke3 {+0.69/19 0s} Nd7 {0.00/21 31s (Nh5)} 35. Kd4
{+0.81/18 23s} Ra5 {0.00/21 0s (Kd8)} 36. Kc3 {+0.85/20
1:43m} Kd6 {+0.06/23 2:20m (Nc5)} 37. Ra1 {+0.84/19 21s
(Rd1+)} Rg5 {+0.28/22 34s} 38. Ra2 {+0.78/20 0s (a5)} Re5
{+0.06/22 8s} 39. b4 {+0.78/20 0s} Re3+ {+0.06/20 0s}
40. Kd4 {+0.74/21 18s} Rb3 {0.00/23 0s} 41. c5+ {+0.74/21
22s} Kc7 {0.00/24 0s (Kc6)} 42. Kc4 {+0.74/20 17s} Rb1
{+0.14/25 13s} 43. Kc3 {+0.83/19 0s (Re2)} Nf6 {+0.07/21
22s (Rc1+)} 44. Rd2 {+0.89/19 14s} Nd7 {0.00/22 11s}
45. Rd5 {+0.89/18 0s} Rc1+ {0.00/20 0s} 46. Kb3 {+0.83/19
15s} Kc6 {+0.14/22 28s} 47. Rd6+ {+0.83/20 0s (Rd4)} Kc7
{0.00/19 0s} 48. Rd4 {+0.83/20 12s (a5)} Rg1 {+0.11/21
1:30m} 49. Rd2 {+0.74/22 0s (Rg4)} Rb1+ {0.00/21 14s}
50. Rb2 {+0.74/21 0s} Ra1 {0.00/19 0s (Rc1)} 51. Re2
{+0.74/19 13s} Kc6 {0.00/23 0s (Rb1+)} 52. Re4 {+0.56/17
24s (Re3)} Rb1+ {0.00/24 26s} 53. Ka2 {+0.70/19 0s (Kc4)}
Rg1 {+0.16/23 2:17m} 54. Re2 {+0.55/20 0s} Nf8 {+0.23/23
18s} 55. Rc2 {+0.55/20 0s} Ne6 {+0.23/21 0s} 56. b5+
{+0.55/19 6s} Kc7 {+0.21/21 0s} 57. Kb2 {+0.55/19 30s (a5)}
Rd1 {+0.05/21 16s} 58. Ka3 {+0.56/18 0s (c6)} Rb1 {0.00/20
15s} 59. Rc4 {+0.56/17 0s (c6)} g6 {0.00/21 1:09m} 60. g4
{+0.57/18 0s (c6)} h5 {0.00/22 16s (h6)} 61. h4 {+0.68/16
13s (gxh5)} f5 {0.00/22 9s} 62. c6 {+0.54/16 0s (gxh5)}
hxg4 {0.00/19 9s} 63. fxg4 {+0.41/17 0s} f4 {0.00/18 0s}
64. Ka2 {+0.68/18 15s} Re1 {+0.06/20 4s} 65. a5 {+0.68/18
1s} f3 {+0.06/20 13s} 66. b6+ {+0.79/18 0s} Kd6 {+0.06/20
0s} 67. b7 {+0.79/19 17s} Kc7 {+0.11/22 0s} 68. Rb4
{+0.79/20 5s} Re2+ {+0.11/22 0s} 69. Ka3 {+0.79/20 5s} Re3+
{+0.11/22 0s} 70. Kb2 {+0.79/20 13s} Re2+ {+0.05/22 0s}
71. Ka1 {+0.79/19 7s (Kc1)} Re1+ {0.00/25 4s} 72. Kb2
{0.00/22 0s} Re2+ {0.00/24 0s} 73. Kb1 {+0.79/18 26s (Kc1)}
Re1+ {0.00/23 12s} 74. Kc2 {0.00/19 0s} Kb8 {0.00/22 0s}
75. a6 {0.00/20 9s} Ra1 {0.00/23 0s} 76. Kd3 {0.00/19 4s
(Re4)} f2 {-0.01/20 9s} 77. a7+ {0.00/20 0s} Rxa7 {-0.01/20
0s} 78. Ke2 {0.00/22 7s} Ra2+ {0.00/22 0s} 79. Kf1 {0.00/22
4s} Rc2 {0.00/23 0s} 80. h5 {0.00/23 12s} Rxc6 {0.00/23 0s
(gxh5)} 81. hxg6 {0.00/22 7s} Nc7 {0.00/24 0s} 82. Kxf2
{0.00/23 6s (g7)} Rxg6 {0.00/25 10s} 83. Kg3 {0.00/24 0s}
Rg8 {0.00/23 0s (Nd5)} 84. Kh4 {0.00/22 5s (Rb6)} Ne6
{0.00/25 15s (Nd5)} 85. Rb6 {0.00/22 5s} Ng5 {0.00/25 0s}
86. Kh5 {0.00/24 3s (Rb5)} Nf3 {0.00/27 14s} 87. Rf6
{0.00/26 0s} Rg5+ {0.00/25 0s} 88. Kh6 {0.00/26 0s} Kxb7
{0.00/23 0s (Rxg4)} 89. Rxf3 {0.00/31 5s} Rxg4 {0.00/30 0s}
90. Kh5 {0.00/59 0s (Rf7+)} Rg7 {0.00/35 9s (Rg8)} 91. Rf6
{0.00/59 0s} Kc7 {0.00/33 0s} 92. Ra6 {0.00/59 0s (Kh6)}
Rg8 {0.00/37 3s} 93. Kh4 {0.00/59 0s (Rg6)} Rg1 {0.00/37
10s (Kd7)} 94. Kh5 {0.00/59 0s} Kd7 {0.00/35 0s} 95. Rb6
{0.00/59 0s (Rf6)} Rg8 {0.00/37 3s (Rh1+)} 96. Kh4 {0.00/59
0s (Rg6)} Rg1 {0.00/37 11s} 97. Kh5 {0.00/59 0s} Rg8
{0.00/38 0s (Rh1+)} 98. Kh4 {0.00/59 0s (Rg6)} Rg1 {0.00/38
4s} 99. Kh5 {0.00/59 0s} 1/2-1/2

Bye
Ingo
Ponder ON rating list: http://www.inwoba.de

User avatar
Matthias Gemuh
Posts: 295
Joined: Wed Jun 09, 2010 2:48 pm
Contact:

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by Matthias Gemuh » Thu Jun 10, 2010 5:52 pm

IWB wrote:Hi all

A thing which I really find very supprising and unsettling is that all the Litos have a ponder hit rate to Rybka 3 (and now R4) which is far beyond my understanding. I am not talking about 'moving fast', I am really talking about ponder hits. This is what is giving these engines an edge over Rybka. This advantage vanishes as soon as you play agains non Rybka engines. What is left then is the "normal" dominance over the opponents as Rybka is showing it.

So for those who can "read" the source code of these Litos and know a bit more about chess programing than I do: What are resonable explanations for such a behavior?

Here is one example game played without a book, ponder on(!) one thread to demonstrate my point. The game is nothing special and a draw anyhow but if you have a look how many moves are replied with 0 seconds on a 30 + 3 time control I have problems to understand that.

Bye
Ingo
It does not surprise me. I expect the ponder hit rate of random-move engines to be near zero, while the ponder hit rate of perfectly strong engines (Elo 9000 ?) should be about 100%.
Rybka and the Litos are much stronger than the field and about equal in strength, so their ponder hit rate has to be quite high. The surprise would be if Stockfish 1.71 does not show similar behaviour.

Matthias.
Aided by engines, GMs can be very strong.
http://www.hylogic.de

Sentinel
Posts: 122
Joined: Thu Jun 10, 2010 12:49 am
Real Name: Milos Stanisavljevic

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by Sentinel » Thu Jun 10, 2010 6:43 pm

Matthias Gemuh wrote:The surprise would be if Stockfish 1.71 does not show similar behaviour.
Actually, if you compare Ingo's list with CCRL, Stockfish performs much better, and the reason for this is (beside better ponder TM of Stockfish compared to Rybka) exactly the high ponder hit rate.

User avatar
IWB
Posts: 195
Joined: Thu Jun 10, 2010 4:10 pm

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by IWB » Thu Jun 10, 2010 8:13 pm

Hi Matthias,
Matthias Gemuh wrote: It does not surprise me. I expect the ponder hit rate of random-move engines to be near zero, while the ponder hit rate of perfectly strong engines (Elo 9000 ?) should be about 100%.
I fully agree here, but Rybka, the Litos or any other top 20 Engine are nearly equaly far off of that perfect 9000 Elo play.
Matthias Gemuh wrote: Rybka and the Litos are much stronger than the field and about equal in strength, so their ponder hit rate has to be quite high. The surprise would be if Stockfish 1.71 does not show similar behaviour.
Matthias.
Yes, that exactly is what I mean. Here is one game, of course without any statistical relevance, where you see the difference (not because of the SF win, but because of the much lower ponder hit rate)


[Event "30 Minutes/Game + 3 Seconds/Move"]
[Site "Engine Match"]
[Date "2010.06.10"]
[Round "1"]
[White "SF 1.7.1 JA 64bit 1T"]
[Black "Fire 1.3 POP x64 BASES 1T"]
[Result "1-0"]

1. e4 {+0.32/24 1:12m} e5 {+0.09/21 44s} 2. Nf3 {+0.24/24
22s} Nc6 {+0.12/20 0s} 3. Bb5 {+0.36/23 54s} Nf6 {+0.22/20
0s (a6)} 4. O-O {+0.40/23 52s (d3)} Bc5 {+0.24/21 1:01m
(Be7)} 5. Nxe5 {+0.48/22 1:01m (Bxc6)} Nxe5 {+0.25/21 41s}
6. d4 {+0.56/24 10s} a6 {+0.32/22 33s (c6)} 7. Be2
{+0.68/23 45s} Bxd4 {+0.28/22 0s} 8. Qxd4 {+0.60/19 2s} d6
{+0.28/21 0s} 9. Rd1 {+0.60/24 47s (Bg5)} Nc6 {+0.14/22
1:28m} 10. Qd3 {+0.72/25 0s} O-O {+0.14/20 0s} 11. c4
{+0.72/22 1:52m (Nc3)} Qe7 {+0.23/19 47s (Nd7)} 12. Nc3
{+0.92/23 1:00m} Nb4 {+0.23/20 0s (Be6)} 13. Qb1 {+1.01/24
54s} Nc6 {+0.26/20 0s (Qe5)} 14. f3 {+0.96/22 57s (Be3)}
Be6 {+0.32/20 37s} 15. Qc2 {+0.96/23 0s} Rfd8 {+0.32/21
1:35m (Nd7)} 16. Be3 {+1.17/23 50s (Nd5)} Ne5 {+0.38/22
1:33m} 17. b3 {+1.21/24 3s (Nd5)} Bd7 {+0.35/19 1:02m
(Re8)} 18. Bg5 {+1.21/25 46s (Rd2)} Bc6 {+0.40/21 28s}
19. Rac1 {+1.17/25 21s (b4)} h6 {+0.34/20 33s} 20. Be3
{+1.21/24 15s} Rac8 {+0.40/20 37s (Re8)} 21. Qd2 {+1.25/22
48s (a4)} Re8 {+0.45/21 1:20m} 22. Bd4 {+1.29/23 10s (a4)}
Rcd8 {+0.48/20 54s (Bd7)} 23. Rc2 {+1.29/22 32s (Nd5)} Bd7
{+0.51/18 33s (Rc8)} 24. Bf2 {+1.25/24 30s} Bc6 {+0.53/20
0s} 25. a4 {+1.25/23 33s (Rdc1)} Ra8 {+0.56/18 33s (Rc8)}
26. Nd5 {+1.17/23 1:49m (Bd4)} Qd8 {+0.59/20 29s} 27. Bd4
{+1.29/22 0s} Nfd7 {+0.69/21 3:55m (Rc8)} 28. Qc3 {+1.41/21
37s} Nf6 {+0.72/21 0s} 29. Ne3 {+1.41/20 40s (Nxf6+)} Qe7
{+0.60/19 18s (Nh5)} 30. Rf1 {+1.61/20 24s (a5)} Rad8
{+0.63/17 44s (Qe6)} 31. c5 {+1.53/19 24s} Qe6 {+0.74/18
10s (Qd7)} 32. Nf5 {+1.57/18 32s} Kh8 {+0.74/17 0s (Nh5)}
33. cxd6 {+1.89/20 30s} cxd6 {+0.93/18 5s} 34. Rcc1
{+1.89/19 25s (Qb2)} Kh7 {+0.68/16 23s (Kg8)} 35. Bb6
{+1.45/17 23s} Rc8 {+0.74/19 0s} 36. Qb4 {+1.21/20 40s} d5
{+0.84/20 0s} 37. Nd4 {+1.25/20 21s} Qd7 {+0.87/20 1s
(Qe7)} 38. f4 {+1.65/21 26s} Ng6 {+0.98/21 29s} 39. e5
{+1.65/22 23s} Ne4 {+0.98/22 41s} 40. Bf3 {+1.53/22 0s
(h3)} Nh4 {+0.81/18 52s (Kg8)} 41. Nxc6 {+1.49/20 20s
(Bh5)} Rxc6 {+0.60/20 23s} 42. Bxe4+ {+1.49/23 0s} dxe4
{+0.60/20 0s} 43. Qxe4+ {+1.45/24 18s} f5 {+0.66/22 0s}
44. Qd4 {+1.41/25 19s} Qe6 {+0.58/23 0s} 45. a5 {+1.61/23
20s} Rxc1 {+0.56/22 0s} 46. Rxc1 {+1.33/23 6s} Qxb3
{+0.56/22 0s} 47. Qd1 {+1.49/24 19s} Qb2 {+0.66/22 0s
(Qxd1+)} 48. Qc2 {+1.45/22 14s} Qxc2 {+0.68/22 5s} 49. Rxc2
{+1.29/23 1s} Kg8 {+0.72/24 27s (g5)} 50. Kf2 {+1.45/23
14s} Ng6 {+0.77/24 16s (Kf7)} 51. Kf3 {+1.37/22 16s (g3)}
h5 {+0.73/24 17s (Nh4+)} 52. g3 {+1.53/23 19s} Re7
{+0.92/24 12s} 53. Rc8+ {+1.81/24 1s (Rc5)} Kf7 {+0.77/22
0s} 54. Rd8 {+1.73/26 12s} Re8 {+0.89/24 0s} 55. Rd7+
{+1.77/25 19s (Rd5)} Re7 {+0.94/26 17s} 56. Rd5 {+1.97/27
0s} Re8 {+1.11/25 1:44m (Re6)} 57. Ke3 {+2.18/25 12s} Nf8
{+1.11/25 0s} 58. Bc5 {+2.26/27 16s} Rc8 {+1.21/25 8s}
59. Bxf8 {+2.18/28 4s} Kxf8 {+1.21/24 0s} 60. Rd7 {+2.30/29
14s} Rc4 {+1.27/25 0s} 61. Rxb7 {+2.38/29 13s} Ra4
{+1.27/26 0s} 62. Ra7 {+2.34/26 23s (h4)} Rxa5 {+1.27/25
11s (Ra3+)} 63. Kd4 {+2.22/27 11s} g6 {+1.27/26 0s (Kg8)}
64. Rc7 {+3.07/25 11s (Rb7)} Rb5 {+1.52/21 24s} 65. Rc6
{+3.27/24 0s} a5 {+1.52/23 9s} 66. Rxg6 {+3.43/24 7s} h4
{+1.52/22 0s} 67. Rf6+ {+3.39/24 13s} Ke7 {+1.50/23 0s}
68. Rxf5 {+3.47/24 10s} a4 {+1.50/23 0s} 69. Rh5 {+3.31/23
10s} Ra5 {+1.36/23 0s} 70. Rh7+ {+3.43/24 18s} Kf8
{+1.36/24 0s (Ke8)} 71. Rc7 {+3.59/26 26s (Rb7)} a3
{+1.51/24 13s} 72. Rc1 {+3.67/27 0s} hxg3 {+1.51/22 0s
(a2)} 73. hxg3 {+3.51/27 27s} a2 {+1.52/28 0s} 74. Ra1
{+4.16/27 10s} Ra4+ {+1.52/29 0s (Ra7)} 75. Kc5 {+4.12/26
8s} Kf7 {+1.58/29 13s (Ke7)} 76. f5 {+4.96/27 15s} Ra5+
{+1.77/29 37s (Re4)} 77. Kd4 {+7.22/25 7s (Kd6)} Ra4+
{+1.77/29 13s (Kg7)} 78. Ke3 {+9.89/22 8s (Kd5)} Ra5
{+1.77/26 10s (Ra3+)} 79. Kf4 {+12.60/24 6s (Ke4)} Ra4+
{+2.24/24 8s} 80. Kg5 {+11.87/23 0s} Ra3 {+2.24/23 0s
(Ra8)} 81. e6+ {+22.14/20 10s (g4)} Ke7 {+5.63/19 16s}
82. g4 {+82.67/21 1s} Ra4 {+5.97/21 10s (Ra8)} 83. Kh5
{+29.93/18 8s} Kf6 {+6.11/22 1s (Kd6)} 84. g5+ {+82.92/30
9s} Kxf5 {+6.12/25 0s} 85. e7 {+83.08/26 0s} Ra8 {+6.15/26
0s} 86. Kh6 {+83.08/29 5s} Re8 {+6.15/28 11s} 87. g6
{+83.13/30 0s} Rxe7 {+6.15/27 0s} 88. g7 {+83.18/28 8s}
Re6+ {+6.18/27 0s} 89. Kh7 {+83.18/30 8s} Re7 {+6.21/27 0s
(Re8)} 90. Kh8 {+83.18/29 8s (Rxa2)} Re2 {+6.41/26 7s
(Re4)} 91. g8=Q {+83.18/29 5s} Rh2+ {+9.35/31 1:15m}
92. Qh7+ {+M18/39 0s} Rxh7+ {+37.31/30 1:25m} 93. Kxh7
{+M17/43 0s} 1-0

I played a couple of thousand games with different Litos, they ALL have a ponder hit rate against Rybka which let them have about double the time left when a critical phase is reached, while they have an equal time management vs the other top engine - and vs Stockfish! (If I would include the Litos in my list, they lose a good part of the advantage if I remove the games vs Rybka)

If you have a Quad computer I can only urge you to make that little experiment. 2 x any Lito, 1 x Rybka3/4 and 1x SF171. Start two games with no book, ponder on and 1 core for each engine.Then have the games run parallel for Rybka-Lito and SF-Lito. You will be very much supprised - I promise!

Still, this "pondertheft" :) is a riddle to me.

Bye
Ingo
Ponder ON rating list: http://www.inwoba.de

Sentinel
Posts: 122
Joined: Thu Jun 10, 2010 12:49 am
Real Name: Milos Stanisavljevic

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by Sentinel » Thu Jun 10, 2010 8:30 pm

IWB wrote:I played a couple of thousand games with different Litos, they ALL have a ponder hit rate against Rybka which let them have about double the time left when a critical phase is reached, while they have an equal time management vs the other top engine - and vs Stockfish! (If I would include the Litos in my list, they lose a good part of the advantage if I remove the games vs Rybka)

If you have a Quad computer I can only urge you to make that little experiment. 2 x any Lito, 1 x Rybka3/4 and 1x SF171. Start two games with no book, ponder on and 1 core for each engine.Then have the games run parallel for Rybka-Lito and SF-Lito. You will be very much supprised - I promise!

Still, this "pondertheft" :) is a riddle to me.
Are you sure it's all due to ponderhit rate and not due to the way ponder is handled. Could you please give the relevant percentages for ponder hit rates for "Litos" as compared to for example SF?

User avatar
IWB
Posts: 195
Joined: Thu Jun 10, 2010 4:10 pm

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by IWB » Thu Jun 10, 2010 9:04 pm

Hi Sentinel
Sentinel wrote: Are you sure it's all due to ponderhit rate and not due to the way ponder is handled. Could you please give the relevant percentages for ponder hit rates for "Litos" as compared to for example SF?
Yes, it is not the way ponderhits are handled. I see a ponder hit of a Lito opponent and an imediately answer by Lito too but that is not what I am talking about as this happens with any opponent.

I do not have an "official" statistic, but it is obvious. Just run the experiment I entioned above and YES again I am sure it is not the way the Litos handle ponderhits.

I counted manualy the two examples I gave here:

Rybka 4 vs Fire = 99 Moves = 53 Ponderhits by Fire = 53.5%
SF171 vs Fire = 93 Moves = 37 Ponderhits by Fire = 39%
(I counted the hits manualy so I might miss one or two hits for both engines, but that difference is in any game.)
In theory one should substract opponent ponderhits from the above numbers, but I am too lazy for that.

As I have written in an erlier post, the Litos win with a huge margin vs the Rybkas, but can not hold that performance BY FAR against any different engine (They performe as an average slightly higher than Rybka 3)

Just play some games of your own - nearly any time control (but longer ones are easier to watch for humans).

Bye
Ingo
Ponder ON rating list: http://www.inwoba.de

Sentinel
Posts: 122
Joined: Thu Jun 10, 2010 12:49 am
Real Name: Milos Stanisavljevic

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by Sentinel » Thu Jun 10, 2010 9:15 pm

IWB wrote:Yes, it is not the way ponderhits are handled. I see a ponder hit of a Lito opponent and an imediately answer by Lito too but that is not what I am talking about as this happens with any opponent.

I do not have an "official" statistic, but it is obvious. Just run the experiment I entioned abouve and YES again I am sure it is not the way the Litos handle ponderhits.

I counted manualy the two examples I gave here:

Rybka 4 vs Fire = 99 Moves = 53 Ponderhits by Fire = 53.5%
SF171 vs Fire = 93 Moves = 37 Ponderhits by Fire = 39%
(I counted the hits manualy so I might miss one or two hits for both engines, but that difference is in any game.)
In theory one should substract opponent ponderhits from the abouve numbers, but I am too lazy for that.

As I have written in an erlier post, the Litos win with a huge margin vs the Rybkas, but can not hold that performance BY FAR against any different engine (They performe as an average slightly higher than Rybka 3)

Just play some games of your own - nearly any time control (but longer ones are easier to watch for humans).
Hi Ingo,
This is interesting. I guess the advantage that you see is due to ponderhit handling (I assume ponderhit rates are reciprocal so if Rybka had the same ponderhit logic it could neutralize the advantage) but this certainly does not explain why ponderhit rate is higher. It can come to similar move selection, but it's hard or almost impossible to tell anything more than that.
There is also an interesting conclusion related to ponderhit implementation coming from your results. This is that fast response ponderhit favors a lot high ponderhit rates, but (there are other data that support this) disfavors low ponderhit rates, so that ponderhit implementations could be improved by simply keeping track of ponderhit rates...

hyatt
Posts: 1242
Joined: Thu Jun 10, 2010 2:13 am
Real Name: Bob Hyatt (Robert M. Hyatt)
Location: University of Alabama at Birmingham
Contact:

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by hyatt » Thu Jun 10, 2010 9:20 pm

I am not sure what I did wrong, but I am posting this again.

ponder hits go up as two programs are closer in overall design. In the "Crafty goes deep" (and others that also tested the same ideas) we found that if you go one ply deeper, a program will change its mind about 15% of the time, or it will stay with the best move from the previous search 85% of the time.

If you play A vs A (same program) then a ponder hit rate of 85% would be pretty common. program A1 searches to depth D to choose its move, and therefore the ponder move is based on a search of depth D-1. When program A2 searches it will go one ply deeper, and change its mind 15% of the time. Staying with the D-1 move 85% of the time and giving A1 85% ponder hits.

If you play such a match without pondering, and count the number of hits that would have occurred had you used pondering, 85% would be expected. High hit rates show lots of similarities between the two programs. You will get pretty high hit rates against different but very strong programs. But in most games, there is not one "best move" in every position. Yet very similar programs will choose the same move, while different but equally strong programs will differ in those positions where there are several equally good alternatives.

Sentinel
Posts: 122
Joined: Thu Jun 10, 2010 12:49 am
Real Name: Milos Stanisavljevic

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by Sentinel » Thu Jun 10, 2010 9:26 pm

hyatt wrote:ponder hits go up as two programs are closer in overall design. In the "Crafty goes deep" (and others that also tested the same ideas) we found that if you go one ply deeper, a program will change its mind about 15% of the time, or it will stay with the best move from the previous search 85% of the time.
If you used Rybka you would get not more than 10%. If you used SF, you would get more than 20%. How reliable are these numbers?
What is the depth where you tried "Crafty goes deep"? My feeling is that the percentage will very with the depth and not necessarily be lower at higher depth. Also I would expect that standard deviation of the given percentage would be more than 5% across various depths.

User avatar
IWB
Posts: 195
Joined: Thu Jun 10, 2010 4:10 pm

Re: Ponder hit rate of the "Litos" to Rybka 3/4

Post by IWB » Thu Jun 10, 2010 9:41 pm

Hi

Actually I dont mind any percentages.

What I recognized was that the Litos perform remakably good vs Rybka. Then I realized that they still win vs the others, but the performance is much lower vs the 18 other engines in my list.
Now I searched for reasons for this remarkable behavior and found the ponder hit rate which let the litos have double the time left sometimes in a game vs Rybka while they rarly have a hughe time bonus left vs the others.

My common sense let me suspect the same as Prof. Hyatt, but as I am not a programmer I can only "suspect" such a thing.

Bye
Ingo
Ponder ON rating list: http://www.inwoba.de

Post Reply