TalkChess.com

Rebel wrote: ↑Mon Jun 03, 2019 3:29 pm I am disappointed, I was wondering if it made sense SF could build an opening book by itself using Multi-PV. First results (analysis time 100ms) looked promising 48.3% against the ProDeo book, 49.4% (analysis time 1000ms), now trying 2000ms but I think I am wasting my time looking at the moves it sometimes produces. Was hoping for a bug that could be fixed but looking at the responses that's unlikely.

You could apply this patch https://github.com/official-stockfish/S ... /pull/2163
which should speedup the MultiPV search to some extent, but I doubt it will noticeably improve the quality of the moves.

Joerg Oster wrote: ↑Mon Jun 03, 2019 4:22 pm

Rebel wrote: ↑Mon Jun 03, 2019 3:29 pm I am disappointed, I was wondering if it made sense SF could build an opening book by itself using Multi-PV. First results (analysis time 100ms) looked promising 48.3% against the ProDeo book, 49.4% (analysis time 1000ms), now trying 2000ms but I think I am wasting my time looking at the moves it sometimes produces. Was hoping for a bug that could be fixed but looking at the responses that's unlikely.

You could apply this patch https://github.com/official-stockfish/S ... /pull/2163
which should speedup the MultiPV search to some extent, but I doubt it will noticeably improve the quality of the moves.

I am not very familiar with Github, where can I find the executable of your patch?

For those interested with SOMU 1.4 - Analyze EPD+++ you can analyze an EPD collection with Multi-PV. Just try it with the example.epd in the EPD folder. The util to make a Polyglot book from the result will be ready likely tomorrow.

http://rebel13.nl/download/utilities.html

Rebel wrote: ↑Mon Jun 03, 2019 9:35 pm

Joerg Oster wrote: ↑Mon Jun 03, 2019 4:22 pm

Rebel wrote: ↑Mon Jun 03, 2019 3:29 pm I am disappointed, I was wondering if it made sense SF could build an opening book by itself using Multi-PV. First results (analysis time 100ms) looked promising 48.3% against the ProDeo book, 49.4% (analysis time 1000ms), now trying 2000ms but I think I am wasting my time looking at the moves it sometimes produces. Was hoping for a bug that could be fixed but looking at the responses that's unlikely.

You could apply this patch https://github.com/official-stockfish/S ... /pull/2163
which should speedup the MultiPV search to some extent, but I doubt it will noticeably improve the quality of the moves.

I am not very familiar with Github, where can I find the executable of your patch?

Sorry, Ed, this is still an open Pull Request under review and has not been merged into master.
I was sure you could do your own compiles ...

Joerg Oster wrote: ↑Tue Jun 04, 2019 11:10 am

Rebel wrote: ↑Mon Jun 03, 2019 9:35 pm

Joerg Oster wrote: ↑Mon Jun 03, 2019 4:22 pm

Rebel wrote: ↑Mon Jun 03, 2019 3:29 pm I am disappointed, I was wondering if it made sense SF could build an opening book by itself using Multi-PV. First results (analysis time 100ms) looked promising 48.3% against the ProDeo book, 49.4% (analysis time 1000ms), now trying 2000ms but I think I am wasting my time looking at the moves it sometimes produces. Was hoping for a bug that could be fixed but looking at the responses that's unlikely.

You could apply this patch https://github.com/official-stockfish/S ... /pull/2163
which should speedup the MultiPV search to some extent, but I doubt it will noticeably improve the quality of the moves.

I am not very familiar with Github, where can I find the executable of your patch?

Sorry, Ed, this is still an open Pull Request under review and has not been merged into master.
I was sure you could do your own compiles ...

Sure, your assumption is right, I just did not expect it works that way.

You are right about the "red", increasing the analysis time to 2000ms and 5000ms thereafter only gave regressions. It's hopeless. I think I will have to switch to Lc0 for the purpose.

Even Leela avoids the direct 3.d4 move and prefers 3.Bb5 instead.

[pgn] [Event "CCC: Testing Updates II"] [Site "?"] [Date "2019.06.02"] [Round "1"] [White "Leelafish"] [Black "Brainfish"] [Result "1-0"] [Variation "Nimzovich-Rossolimo attack (without ...d6)"] [TimeControl "600+2"] [Termination "normal"] [PlyCount "259"] 1. e4 c5 2. Nf3 Nc6 3. Bb5 e6 4. Bxc6 bxc6 5. d3 Ne7 6. h4 h5 7. e5 f6 8. Qe2 Qc7 9. b3 fxe5 10. Bb2 d6 11. Nbd2 Nf5 12. O-O-O Be7 13. Kb1 Bf6 14. Ne4 Qe7 15. Qd2 Bb7 16. Qa5 Qd8 17. Qd2 Qe7 18. Qa5 Qd8 19. Qe1 Qd7 20. Nfg5 O-O-O 21. f3 Rdf8 22. g4 Nd4 23. c3 Nb5 24. Nxf6 gxf6 25. Ne4 f5 26. Nd2 Nc7 27. c4 hxg4 28. fxg4 Rfg8 29. g5 Qe8 30. Qe3 Qg6 31. Rh2 Na6 32. Rg1 Rh7 33. Qe2 Qh5 34. Qe3 Rhg7 35. a3 Kb8 36. Nf1 Nc7 37. Qf2 Ne8 38. Ng3 Qh8 39. Bc1 Rh7 40. Rgh1 f4 41. Ne4 Rh5 42. Bb2 Qh7 43. Qf3 Ka8 44. Rg1 Qf7 45. Ka2 Qg6 46. a4 Rgh8 47. Rg4 Rf8 48. a5 a6 49. Bc3 Rg8 50. Rh1 Ka7 51. Rh2 Ka8 52. Rh1 Ka7 53. Qf2 Ka8 54. Rhg1 Rh7 55. R4g2 Qf5 56. Rh2 Qg6 57. Qf3 Rh5 58. Qf2 Qh7 59. Qf3 Qf7 60. Be1 Rgh8 61. g6 Qe7 62. Ng5 Ng7 63. Nf7 Rf8 64. Rg5 Rxg5 65. Nxg5 Rf5 66. Bc3 Kb8 67. Be1 Ka8 68. Qg4 Bc8 69. Nf7 Rh5 70. Rb2 Rf5 71. Qf3 Bb7 72. Qg4 Qd7 73. Ng5 Qe7 74. Rg2 f3 75. Nxf3 Rf4 76. Qh3 Qf6 77. Nd2 Qf5 78. Qxf5 exf5 79. Rg5 Bc8 80. h5 Nxh5 81. Rxh5 Rg4 82. Rh6 Kb7 83. Kb2 Be6 84. g7 Bg8 85. Rxd6 Rxg7 86. Kc3 Kc7 87. Rf6 Rf7 88. Rh6 e4 89. dxe4 f4 90. Bf2 Rh7 91. Rf6 Rh3+ 92. Kc2 Rh5 93. Rxf4 Be6 94. Rf8 Rg5 95. Be3 Rh5 96. Bf4+ Kd7 97. Ra8 Rh4 98. Be3 Bg4 99. Ra7+ Kc8 100. Rxa6 Kc7 101. Bf4+ Kb7 102. Rb6+ Ka8 103. Rxc6 Rh5 104. Bd6 Bd7 105. Rc7 Rh1 106. Bxc5 Bg4 107. b4 Rh8 108. b5 Bc8 109. a6 Rg8 110. Bd6 Bxa6 111. bxa6 Rg7 112. Rc8+ Ka7 113. Bc5+ Kxa6 114. Ra8+ Kb7 115. Ra7+ Kc6 116. Rxg7 Kxc5 117. e5 Kd4 118. e6 Ke5 119. e7 Kf6 120. e8=Q Kxg7 121. c5 Kf6 122. c6 Kg5 123. c7 Kf4 124. c8=Q Kg5 125. Qf8 Kh4 126. Qg4+ Kxg4 127. Ne4 Kh5 128. Qg7 Kh4 129. Qg5+ Kh3 130. Qg3# 1-0 [/pgn]

Unusual game, giving up pawns for free in openning and getting unknown compensation then went on to win.


edit: Ok I looked at it again, and saw the weird compensation seems to be the black N being misplaced and can't reach an ideal outpost until the end of the game. But really, that's too deep a compensation for the gambit pawn given at move nine and realizing the compensation at move 30. .

MikeGL wrote: ↑Wed Jun 05, 2019 2:52 pm Even Leela avoids the direct 3.d4 move and prefers 3.Bb5 instead.

But Leela, like most average engines, will have 3.d4 in their list with Multi-PV=3.

MikeGL wrote: ↑Wed Jun 05, 2019 2:52 pm Even Leela avoids the direct 3.d4 move and prefers 3.Bb5 instead.

[Event "CCC: Testing Updates II"][Date "2019.06.02"][Round "1"]
[White "Leelafish"][Black "Brainfish"]
[Result "1-0"]
[Variation "Nimzovich-Rossolimo attack (without ...d6)"]

Unusual game, giving up pawns for free in openning and getting unknown compensation then went on to win.

edit: Ok I looked at it again, and saw the weird compensation seems to be ...

Well, we don't know those were good moves just because it went on to win the game. What was the TC?

jp wrote: ↑Wed Jun 05, 2019 7:23 pm

MikeGL wrote: ↑Wed Jun 05, 2019 2:52 pm Even Leela avoids the direct 3.d4 move and prefers 3.Bb5 instead.

[Event "CCC: Testing Updates II"][Date "2019.06.02"][Round "1"]
[White "Leelafish"][Black "Brainfish"]
[Result "1-0"]
[Variation "Nimzovich-Rossolimo attack (without ...d6)"]

Unusual game, giving up pawns for free in openning and getting unknown compensation then went on to win.

edit: Ok I looked at it again, and saw the weird compensation seems to be ...

Well, we don't know those were good moves just because it went on to win the game. What was the TC?

[Event "CCC: Testing Updates II"]
[Site "?"]
[Date "2019.06.02"]
[Round "1"]
[White "Leelafish"]
[Black "Brainfish"]
[Result "1-0"]
[ECO "B30"]
[GameDuration "00:30:38"]
[GameEndTime "2019-06-02T02:48:50.956 PDT"]
[GameStartTime "2019-06-02T02:18:12.260 PDT"]
[Opening "Sicilian"]
[PlyCount "259"]
[TimeControl "600+2"]
[Variation "Nimzovich-Rossolimo attack (without ...d6)"]

so 10 mins + 2 seconds increment.
Sorry, the PGN header was cleaned up by my GUI during copy paste.

Rebel wrote: ↑Mon Jun 03, 2019 3:29 pm I think I am wasting my time looking at the moves it sometimes produces. Was hoping for a bug that could be fixed but looking at the responses that's unlikely.

There's a bug, but it's not on the engine.

This situation happened (note, this is not about Stockfish, all engines suck like this, unless they get helped by theory, like ProDeo or Shredder <12, or Eman's Experience File) because engine testers chose to play with generic books that take over the opening and play reasonable moves for the engine, and when the engine has to play it already skipped the problem entirely.

Engine developers have had no reason whatsoever to get better at the opening, and self-play has had a similar resolution. The elo chase is real, and even Komodo, an engine that has been created with "analysis first" on mind, doesn't behave better in the opening (heck, it doesn't behave better in the early middle game either.)

InfinityChess's Chess Engine Master tournament has shown that Leela isn't good at the openings either, because people are gaining much better advantage by using books, and it seems a book is good on its own and it doesn't matter if Stockfish or Leela play when out of book, and Leela doesn't have the time to find a better move than the one the book'd have played.