Pro Deo 1.5

[ed]

After intensive research I finally found an successful way to improve the playing strength of Pro Deo using a chess tree. Since the format of the chess tree (called: EOC) is an open source every chess programmer can add the algorithm to his/her own engine and gain some elo points without too much of a hassle.

Download and explanations: http://www.top-5000.nl/prodeo.htm

How to make and access EOC: http://www.top-5000.nl/eoc.htm

Technical issues:

1) The EOC source code doesn't contain the Pro Deo formula used to calculate the bonus/penalty for each move found in the EOC chess tree, if there is interest I would happy to provide it.

2) New in Pro Deo 1.5 is also the option the automatically update the EOC chess tree while playing its games, it seems to help too. It's controlled by the following parameter in a personality.

[EOC Learner = on] * off|on

On request I can provide the source code of this routine too.

Observations.......

The last months I have tried several EOC databases, using it as an opening book gave terrible results, using a giant 100 million positions (!!) chess tree (unfortunately) was unsuccessful too.

Then I made an EOC from about 120,000 quality comp-comp games between the 20 best chess engines, added about 30,000 human 2500+ elo games and the first positive result was a fact.

But the obvious best result gave an EOC database made by 6700 (quality!) games Pro Deo played itself. If you think about it for a while it's not such a big surprise as the created chess tree reflects the total history of the opening and middle game Pro Deo has played, avoiding bad lines and stimulating good lines. As such the system more or less automatically functions as a learner, especially since the EOC database is updated when the result of a game becomes known, score: +5.00 cq -5.00

This was a fun project and I am open to offer my help to anyone who wants to add this technique to his/her engine.

Ed

[Tony]

Hi Ed,

yes I'm interested in your formula to create a score from the winning percentage, play amount (maybe elo).

When you used the big eoc tree, did you try combining it with a 10 ply search (including search score offset by above formula) ?

Cheers,

Tony

[ed]

Hi Ed,

yes I'm interested in your formula to create a score from the winning percentage, play amount (maybe elo).

OK, will update the source soon.

When you used the big eoc tree, did you try combining it with a 10 ply search (including search score offset by above formula) ?

I don't understand the question. What I do is this: for each root move I call the EOC database once, calculate the bonus/penalty and store the result in a table. Each new iteration this score is added to "incremental score" of the root move.

Not sure if I express myself clearly now.

Ed

Cheers,

Tony

[mjlef]

Hmmm... I wonder if that is what all that data is in Rybka? Bonuses for lines it tends to win more in.

[CRoberson]

Hi Ed,

This sounds great. I've put quite a bit of thought into this and had
several discussions with a colleague who is into Bayesian trees.
Effectively, this is a Bayesian tree. My debate is: while using simple
counting statistics works, it may not be the most effective.

I believe the quality of the effort needs to be reflected -- more
bonus for more important wins and more penalty for bigger losses.

With that in mind, why not give more penalty for a loss/draw against a
weaker opponent and more bonus for a win/draw against a stronger
opponent. The values should be scaled evenly across the range of
strong vs weaker.

Charles

[Uri Blass]

My tests are without book so at this point of time this is not relevant for me(I may be interested in it later) but I tried to run prodeo(with no book) against movei under chessbase gui and afrer the result was 19-19 prodeo simply crashed

I wonder if you plan to fix that bug(otherwise I will have to choose different opponents for movei because I do not like to lose cpu time because of crashes)

Note that I had no single case of crashes in games against fruit2.0
but some programs like prodeo1.2 and zappa1.1 are simply unstable and can crash after long games.


[Event "URI-AMD, Blitz:1'+1""]
[Site "URI-AMD"]
[Date "2007.06.07"]
[Round "38"]
[White "Movei00_8_423"]
[Black "Pro Deo 1.2"]
[Result "1/2-1/2"]
[ECO "B08"]
[Annotator "0.84;0.37"]
[PlyCount "405"]
[TimeControl "60+1"]

{W=10.9 ply; 535kN/s B=8.3 ply; 1,060kN/s} 1. e4 d6 2. d4 Nf6 3. Nc3 g6 4. Nf3
Bg7 5. Be2 O-O 6. O-O Bg4 {Both last book move} 7. Be3 {0.84/9 2} Nbd7 {
(e7-e6) 0.37/10 2} 8. d5 {(h2-h3) 0.68/9 3} a6 {(Bg4xf3) 0.17/10 3} 9. Kh1 {
(h2-h3) 0.76/9 2} Nb6 {(Bg4xf3) 0.08/9 3} 10. Bd4 {(Nf3-d4) 0.77/8 3} Bh6 {
(Bg4-d7) 0.06/9 2} 11. Qd3 {(Nf3-d2) 0.69/10 6} c5 {(Bh6-f4) 0.05/10 3} 12.
dxc6 {0.85/10 2} bxc6 {0.24/11 6} 13. Bxf6 {1.05/9 2} exf6 {0.12/6 0} 14. Rfd1
{(Ra1-d1) 1.00/10 2} Nc8 {(Bg4xf3) 0.16/10 2} 15. Nd4 {0.90/9 2} Bd7 {0.19/11 6
} 16. Rab1 {(Qd3-g3) 0.85/8 1} Re8 {0.13/10 2} 17. b4 {(g2-g3) 0.71/9 3} Qc7 {
(Nc8-b6) 0.10/9 2} 18. Qf3 {(Qd3-g3) 0.76/9 3} Bg7 {0.14/9 2} 19. Qf4 {0.56/9 3
} Nb6 {0.10/9 2} 20. Qg3 {(Rb1-c1) 0.58/9 4} f5 {0.08/9 2} 21. exf5 {0.27/10 2}
Be5 {-0.17/9 2} 22. Qh3 {(f2-f4) 0.08/9 1} Bxd4 {(c6-c5) -0.19/8 1} 23. Rxd4 {
0.23/10 2} Bxf5 {-0.19/10 2} 24. Qg3 {0.20/10 2} Re5 {-0.27/10 2} 25. Rd2 {
(f2-f4) 0.15/9 4} Qe7 {(Ra8-e8) -0.26/9 2} 26. Qf3 {(Rb1-d1) -0.12/9 2} d5 {
-0.60/9 2} 27. a3 {(h2-h3) -0.50/9 2} Qg5 {(Ra8-e8) -1.00/9 1} 28. Rd4 {
(Qf3-g3) -0.93/9 1} Bxc2 {-1.11/9 2} 29. h4 {(Rb1-a1) -1.14/9 2} Qe7 {
-1.39/10 2} 30. Rb2 {-1.19/9 1} Bf5 {-1.42/9 1} 31. Kg1 {(Kh1-h2) -1.32/8 1}
Re8 {-1.55/9 1} 32. h5 {(Rb2-d2) -1.43/9 2} Nc4 {(Nb6-a4) -1.63/8 1} 33. Ra2 {
-1.23/8 1} Qa7 {(Qe7-g5) -1.58/8 2} 34. Rf4 {(Rd4-d1) -1.49/8 2} Nd6 {-1.56/9 2
} 35. Rd2 {(h5xg6) -1.40/9 3} Nb5 {(g6xh5) -1.58/8 2} 36. Nxb5 {-1.59/9 3} axb5
{-1.51/5 0} 37. Kf1 {(h5xg6) -1.69/9 2} c5 {(Bf5-e4) -1.55/8 2} 38. bxc5 {
(h5xg6) -1.75/8 1} Qxc5 {-1.47/5 0} 39. g3 {-2.07/9 2} Bh3+ {-1.95/8 1} 40. Kg1
{-1.87/10 2} Rxh5 {-1.89/9 1} 41. Rfd4 {-1.74/9 1} Rf5 {-1.85/9 1} 42. Qd3 {
-1.58/9 2} Rfe5 {-2.08/9 1} 43. Qxb5 {(Rd4-b4) -1.61/9 2} Qe7 {
(Qc5xa3) -1.61/9 1} 44. Rh4 {(Be2-f3) -1.74/9 1} Rxe2 {(d5-d4) -1.75/5 0} 45.
Qxe2 {-1.52/12 1} Qxe2 {-1.82/5 0} 46. Rxe2 {-1.36/12 1} Rxe2 {-1.55/5 0} 47.
Rxh3 {-1.38/11 1} Re4 {(Re2-e1+) -1.68/12 1} 48. Rh4 {(Kg1-g2) -0.99/10 1} Rxh4
{-1.83/12 1} 49. gxh4 {-0.69/11 1} Kf8 {-1.82/16 2} 50. f4 {(Kg1-f1) -0.64/13 3
} Ke7 {-1.84/16 1} 51. Kf2 {-0.64/13 2} Kd6 {-1.89/15 1} 52. a4 {
(Kf2-e3) -0.74/13 3} Kc5 {-1.79/15 2} 53. Ke3 {-0.66/13 1} f5 {
(f7-f6) -1.57/15 2} 54. a5 {-0.51/14 1} Kb5 {-0.25/15 1} 55. Kd4 {-0.44/15 1}
Kxa5 {-0.15/12 0} 56. Kxd5 {0.00/14 1} Kb4 {(Ka5-b5) -0.26/18 1} 57. Ke5 {
(Kd5-e6) 0.00/15 1} Kc4 {-0.24/18 1} 58. Kf6 {0.00/16 1} Kd5 {-0.17/19 2} 59.
Kg7 {0.00/18 1} Ke4 {0.00/20 3} 60. Kxh7 {0.00/21 1} Kxf4 {-0.22/5 0} 61. Kg7 {
(Kh7xg6) 0.00/1 0} g5 {-1.53/15 1} 62. h5 {0.00/1 0} g4 {-1.52/14 1} 63. h6 {
0.00/1 0} g3 {-1.53/13 1} 64. h7 {0.00/1 0} g2 {-1.58/14 3} 65. h8=Q {0.00/1 0}
g1=Q+ {-1.36/12 2} 66. Kf7 {0.00/1 0} Qa7+ {-1.58/8 2} 67. Kg6 {
(Kf7-e6) 0.00/1 0} Qa6+ {-1.71/8 1} 68. Kh5 {(Kg6-f7) 0.00/1 0} Qe2+ {-1.61/9 1
} 69. Kg6 {0.00/1 0} Qg2+ {-1.63/9 2} 70. Kf7 {0.00/1 0} Qd5+ {-1.66/9 2} 71.
Kg6 {0.00/1 0} Kg3 {-1.48/9 3} 72. Qc3+ {0.00/1 0} Kg4 {-1.44/9 2} 73. Qc8 {
0.00/1 0} Qd6+ {-1.44/9 2} 74. Kf7 {0.00/1 0} Qh6 {-1.43/9 3} 75. Qd8 {
(Qc8-c2) 0.00/1 0} Qe3 {-1.42/9 2} 76. Qc8 {(Qd8-d7) 0.00/1 0} Qd2 {-1.41/9 2}
77. Qe6 {(Kf7-f6) 0.00/1 0} Qd3 {-1.35/9 2} 78. Qg6+ {0.00/1 0} Kf4 {-1.39/8 2}
79. Qh5 {(Qg6-h6+) 0.00/1 0} Qd7+ {-1.39/8 2} 80. Kf8 {(Kf7-f6) 0.00/1 0} Qd5 {
-1.35/8 1} 81. Qg6 {(Kf8-e7) 0.00/1 0} Qe4 {-1.58/8 1} 82. Qh6+ {0.00/1 0} Kg3
{-1.57/8 2} 83. Qg7+ {0.00/1 0} Kf2 {-1.60/8 2} 84. Qb2+ {0.00/1 0} Kf3 {
-1.53/5 0} 85. Qc3+ {0.00/1 0} Qe3 {-1.48/8 1} 86. Qc6+ {0.00/1 0} Kg4 {
-1.50/9 2} 87. Qg6+ {0.00/1 0} Qg5 {-1.46/5 0} 88. Qe6 {0.00/1 0} Kf3 {
-1.44/5 0} 89. Qd5+ {0.00/1 0} Kg3 {-1.51/8 3} 90. Qe5+ {0.00/1 0} Qf4 {
-1.44/5 0} 91. Qc3+ {0.00/1 0} Kg4 {-1.42/5 0} 92. Kf7 {0.00/1 0} Qe4 {
-1.42/5 0} 93. Qg7+ {0.00/1 0} Kf4 {-1.43/8 2} 94. Qh6+ {0.00/1 0} Kg3 {
-1.45/5 0} 95. Qg7+ {0.00/1 0} Kf2 {-1.43/8 2} 96. Qb2+ {0.00/1 0} Qe2 {
-1.41/5 0} 97. Qc1 {0.00/1 0} Kf3 {-1.40/8 1} 98. Qh1+ {(Qc1-c3+) 0.00/1 0} Qg2
{-1.42/5 0} 99. Qd1+ {0.00/1 0} Ke3 {-1.35/8 2} 100. Qb3+ {0.00/1 0} Kf4 {
-1.33/5 0} 101. Qd1 {(Qb3-e6) 0.00/1 0} Qe4 {-1.34/8 2} 102. Qh5 {
(Kf7-f6) 0.00/1 0} Ke3 {-1.40/9 3} 103. Qd1 {0.00/1 0} f4 {-1.49/5 0} 104. Qe1+
{0.00/1 0} Kd4 {-1.46/5 0} 105. Qa1+ {(Qe1-b4+) 0.00/1 0} Kc4 {-1.48/5 0} 106.
Qa2+ {(Qa1-a4+) 0.00/1 0} Kc3 {-2.04/8 4} 107. Kf6 {(Qa2-a1+) 0.00/1 0} f3 {
-1.60/5 0} 108. Kg5 {0.00/1 0} Qe3+ {-3.07/5 0} 109. Kh4 {(Kg5-h5) 0.00/1 0} f2
{-3.39/5 0} 110. Qf7 {0.00/1 0} Kd3 {-3.42/5 0} 111. Qb3+ {0.00/1 0} Ke2 {
-3.03/8 1} 112. Qb5+ {0.00/1 0} Kf3 {-3.10/8 2} 113. Qh5+ {0.00/1 0} Ke4 {
-3.03/5 0} 114. Qg6+ {(Qh5-h7+) 0.00/1 0} Kd5 {-3.10/8 2} 115. Qf7+ {0.00/1 0}
Kd6 {-3.12/5 0} 116. Qg6+ {0.00/1 0} Kc7 {-3.13/8 2} 117. Qf7+ {
(Qg6-h7+) 0.00/1 0} Kc6 {-3.14/5 0} 118. Qg6+ {(Qf7-c4+) 0.00/1 0} Kb7 {
-2.88/8 1} 119. Qf7+ {(Qg6-b1+) 0.00/1 0} Kb6 {-3.12/8 2} 120. Qg6+ {0.00/1 0}
Ka5 {-3.11/5 0} 121. Qf7 {(Qg6-f5+) 0.00/1 0} Qe2 {-3.06/5 0} 122. Qd5+ {
0.00/1 0} Kb4 {-3.06/9 2} 123. Qb7+ {(Qd5-d6+) 0.00/1 0} Kc5 {-3.12/5 0} 124.
Qc8+ {(Qb7-c7+) 0.00/1 0} Kd4 {-2.68/9 4} 125. Qd7+ {(Qc8-d8+) 0.00/1 0} Kc3 {
-3.16/5 0} 126. Qg7+ {0.00/1 0} Kd3 {-3.25/5 0} 127. Qd7+ {0.00/1 0} Kc4 {
-3.11/5 0} 128. Qc8+ {(Qd7-f7+) 0.00/1 0} Kb3 {-3.12/9 1} 129. Qb7+ {
(Qc8-b8+) 0.00/1 0} Kc2 {-3.09/9 2} 130. Qc6+ {(Qb7-c8+) 0.00/1 0} Kd2 {
-3.03/5 0} 131. Qd5+ {0.00/1 0} Kc3 {-3.04/5 0} 132. Qa5+ {0.00/1 0} Kb3 {
-1.55/9 2} 133. Qb6+ {0.00/1 0} Ka3 {-3.03/5 0} 134. Qa7+ {(Qb6-a5+) 0.00/1 0}
Kb2 {-0.30/9 2} 135. Qb6+ {(Qa7-d4+) 0.00/1 0} Kc1 {-2.95/5 0} 136. Qc5+ {
0.00/1 0} Kd1 {-2.64/8 1} 137. Qd4+ {0.00/1 0} Ke1 {-2.94/5 0} 138. Qc3+ {
(Qd4-a1+) 0.00/1 0} Qd2 {-0.60/9 2} 139. Qa1+ {(Qc3-e5+) 0.00/1 0} Ke2 {
-2.59/5 0} 140. Qe5+ {0.00/1 0} Kd1 {-0.34/8 2} 141. Qa1+ {0.00/1 0} Qc1 {
-0.21/5 0} 142. Qd4+ {0.00/1 0} Ke2 {-0.14/9 1} 143. Qe5+ {(Qd4-g4+) 0.00/1 0}
Kf1 {-3.04/9 3} 144. Qd5 {(Qe5-e4) 0.00/1 0} Qf4+ {-3.10/5 0} 145. Kh3 {
0.00/1 0} Ke2 {-3.16/5 0} 146. Qb5+ {0.00/1 0} Kf3 {-0.07/8 1} 147. Qd3+ {
0.00/1 0} Qe3 {-0.11/3 0} 148. Qf5+ {0.00/1 0} Ke2+ {0.00/9 2} 149. Kg2 {
0.00/1 0} Ke1 {-0.24/5 0} 150. Qb1+ {0.00/1 0} Kd2 {0.00/10 2} 151. Qa2+ {
(Qb1-b2+) 0.00/1 0} Kc3 {0.00/5 0} 152. Qd5 {(Qa2-a3+) 0.00/1 0} f1=Q+ {
0.00/10 1} 153. Kxf1 {0.00/1 0} Kb4 {0.00/10 2} 154. Qc6 {(Qd5-d6+) 0.00/1 0}
Qd2 {-0.14/5 0} 155. Qb7+ {(Qc6-g2) 0.00/1 0} Kc5 {-0.17/5 0} 156. Qa8 {
(Qb7-g2) 0.00/1 0} Qb2 {0.00/9 5} 157. Qe4 {(Qa8-a5+) 0.00/1 0} Kd6 {-0.13/5 0}
158. Qa8 {(Kf1-e1) 0.00/1 0} Ke6 {-0.16/5 0} 159. Qc6+ {(Qa8-e4+) 0.00/1 0} Ke5
{0.00/5 0} 160. Qa8 {(Qc6-c5+) 0.00/1 0} Kf4 {-0.20/5 0} 161. Qc6 {
(Qa8-d5) 0.00/1 0} Qa2 {-0.19/5 0} 162. Qb7 {(Qc6-g6) 0.00/1 0} Qc4+ {-0.10/8 2
} 163. Kg1 {(Kf1-e1) 0.00/1 0} Qc2 {-0.21/5 0} 164. Qa8 {(Qb7-d5) 0.00/1 0} Qd2
{-0.19/8 1} 165. Qb7 {(Qa8-g2) 0.00/1 0} Qe1+ {-0.11/8 2} 166. Kg2 {0.00/1 0}
Qe2+ {-0.20/5 0} 167. Kg1 {(Kg2-h3) 0.00/1 0} Qe5 {-0.07/8 1} 168. Qa8 {
(Kg1-f2) 0.00/1 0} Qe3+ {-0.14/8 2} 169. Kh1 {(Kg1-f1) 0.00/1 0} Qd2 {-0.23/5 0
} 170. Qb7 {(Qa8-g2) 0.00/1 0} Qc2 {-0.17/8 1} 171. Qa8 {(Kh1-g1) 0.00/1 0} Ke5
{-0.28/5 0} 172. Qb7 {(Kh1-g1) 0.00/1 0} Qc4 {-0.08/9 3} 173. Qa8 {
(Qb7-e7+) 0.00/1 0} Kf4 {-0.24/5 0} 174. Qb7 {(Kh1-h2) 0.00/1 0} Qf1+ {
-0.23/5 0} 175. Kh2 {0.00/1 0} Qf2+ {-0.10/9 2} 176. Kh1 {(Qb7-g2) 0.00/1 0}
Qa2 {-0.23/5 0} 177. Qc6 {(Qb7-g2) 0.00/1 0} Ke5 {-0.26/5 0} 178. Qb7 {
(Kh1-g1) 0.00/1 0} Qe6 {-0.06/9 2} 179. Qa8 {(Kh1-g2) 0.00/1 0} Kf4 {-0.25/5 0}
180. Qb7 {(Kh1-h2) 0.00/1 0} Ke3 {-0.18/8 1} 181. Qa8 {(Kh1-g2) 0.00/1 0} Qh3+
{-0.24/5 0} 182. Kg1 {0.00/1 0} Qg3+ {-0.13/9 3} 183. Kh1 {(Qa8-g2) 0.00/1 0}
Qg5 {-0.22/5 0} 184. Qb7 {(Kh1-h2) 0.00/1 0} Kf4 {-0.23/5 0} 185. Qa8 {
(Qb7-b4+) 0.00/1 0} Qb5 {-0.14/9 2} 186. Qg2 {(Qa8-f8+) 0.00/1 0} Qd3 {
-0.23/5 0} 187. Qc6 {(Kh1-h2) 0.00/1 0} Qf1+ {-0.21/5 0} 188. Kh2 {0.00/1 0}
Qf2+ {-0.08/9 2} 189. Kh1 {0.00/1 0} Qa2 {-0.23/5 0} 190. Qb7 {
(Qc6-g2) 0.00/1 0} Qa5 {-0.07/9 4} 191. Qc6 {(Kh1-g2) 0.00/1 0} Qe1+ {-0.18/5 0
} 192. Kh2 {(Kh1-g2) 0.00/1 0} Qg3+ {-0.21/5 0} 193. Kh1 {0.00/1 0} Qb3 {
-0.17/5 0} 194. Qa8 {(Kh1-g2) 0.00/1 0} Qb2 {-0.23/5 0} 195. Qc6 {
(Qa8-g2) 0.00/1 0} Qb8 {0.00/8 2} 196. Qd5 {(Qc6-f6+) 0.00/1 0} Qc8 {-0.17/5 0}
197. Qg2 {(Kh1-g2) 0.00/1 0} Qa6 {0.00/11 1} 198. Qd5 {(Qg2-f2+) 0.00/1 0} Qg6
{-0.15/5 0} 199. Qb7 {(Qd5-d4+) 0.00/1 0} Qf6 {0.00/13 2} 200. Qa8 {
(Kh1-g2) 0.00/1 0} Qd6 {0.00/5 0} 201. Qb7 {(Kh1-g2) 0.00/1 0} Qf8 {0.00/15 2}
202. Qc6 {(Kh1-g2) 0.00/1 0} Ke5 {0.00/5 0} 203. Qb7 {Draw accepted 0.00/1 0}
1/2-1/2

[ed]

Hi Ed,

yes I'm interested in your formula to create a score from the winning percentage, play amount (maybe elo).

When you used the big eoc tree, did you try combining it with a 10 ply search (including search score offset by above formula) ?

Cheers,

Tony

Updated the source with the formula. Because of technical reasons it's not 100% the same as in Pro Deo 1.5 but watch the code and you get the drift, nothing special and arguable too. I am sure the formula can be improved.

http://www.top-5000.nl/eoc.htm

Ed

[ed]

Hi Uri,

Pro Deo Frequently Asked Questions

When you are playing eng-eng matches or auto232 make sure you limit the maximum number of moves else PRO DEO might play funny moves or worse crash.

For ChessBase this means: Move Limit = 240
For LokaSoft: Max Moves in Game = 160

http://www.top-5000.nl/faq.htm

Ed

[ed]

Hi Ed,

This sounds great. I've put quite a bit of thought into this and had
several discussions with a colleague who is into Bayesian trees.
Effectively, this is a Bayesian tree. My debate is: while using simple
counting statistics works, it may not be the most effective.

I believe the quality of the effort needs to be reflected -- more
bonus for more important wins and more penalty for bigger losses.

With that in mind, why not give more penalty for a loss/draw against a
weaker opponent and more bonus for a win/draw against a stronger
opponent. The values should be scaled evenly across the range of
strong vs weaker.

Charles

Hi Charles,

Interesting thoughts indeed. Maybe I should do some experiments building the EOC chess tree based on the opponent's elo rating.

And I thought the project was finished....

Ed

[Tony Thomas]

I can verify that Zappa got problems as a UCI engine. However, under arena, Zappa 1.1 hasnt crashed a single time as a WB engine.