Leela Network #31431 Beats Stockfish at 1 minute per move!

[Harbaugh]

Time control: 10 moves in 10 min, repeating

Leela v0.20.0-dev Network #31431 Hardware:  2x Tesla V100's (approx 30-50 Knps) 6-man TB

Stockfish 9 x64 Popcnt Hardware: AMD FX-8350 4 cores 4000MHz 24GB (approx 5 Mnps)

2-ply opening book:  1. e4 d5

Exciting to see last weekend how the new networks were pressing against Stockfish , and today she's converting to wins!

Leela is growing stronger by the day & I'll be testing long time controls all weekend updating Leela's net# as I go. After a few more wins I'll probably setup a stockfish beast with 96 cores etc to see if she can handle Stockfish at its best.

Funny how Leela tries to give away a piece at the end so she can get to 6-men and use her tablebase. Also her two underpromotions. But watching her try to mate a bare king is like watching paint dry. I have both WDL and DTZ installed and Leela detected them both at startup, but it looks like she is only evaluating with the WDL.  I'm attaching Arena debug info if anyone wants to decipher what is going on with that.

---------

Short PGN as follows:  (Long PGN attached)

--------

[Date "2018.11.21"]

[White "Remote Lc0"]

[Black "Local Stockfish"]

[Result "1-0"]

[ECO "B01"]

[Opening "Scandinavian"]

[Time "22:32:23"]

[Variation "Marshall, 4.c4 Nb6 5.Nf3 g6 6.Nc3 Bg7"]

[TimeControl "10/600:10/600:10/600"]

1. e4 d5 2. exd5 Nf6 3. d4 Nxd5 4. Nf3 g6 5. c4 Nb6 6. Nc3 Bg7 7. c5 Nd5 8.

Bc4 Nxc3 9. bxc3 O-O 10. Bg5 Nc6 11. O-O Na5 12. Bd3 Re8 13. Re1 c6 14. Re3

f6 15. Bh4 b6 16. Qc2 Bg4 17. Rae1 Bxf3 18. gxf3 bxc5 19. Re6 Kf7 20. f4

cxd4 21. f5 g5 22. Bg3 Kg8 23. Be2 d3 24. Bxd3 Bf8 25. R1e4 Nb7 26. Rd4 Nd6

27. h4 g4 28. c4 c5 29. Rxg4+ Kh8 30. Bxd6 exd6 31. Be4 Rc8 32. Bd5 Bg7 33.

Qe4 Rxe6 34. fxe6 Qe7 35. Qf5 Rg8 36. Kh2 Bh6 37. Rg3 Rxg3 38. fxg3 Bc1 39.

Be4 Bh6 40. Qxh7+ Qxh7 41. Bxh7 Kg7 42. Kh3 Bd2 43. Be4 Be1 44. h5 Kf8 45.

h6 f5 46. Bxf5 Ba5 47. Kg4 Bc3 48. Kh5 a5 49. g4 d5 50. cxd5 Ke7 51. g5 Kd6

52. g6 Bf6 53. g7 c4 54. g8=Q c3 55. Bc2 Ke5 56. h7 a4 57. h8=B Bxh8 58.

Qxh8+ Kd6 59. Qxc3 a3 60. Qxa3+ Ke5 61. Qf8 Kd4 62. Bd3 Kc3 63. Qb4+ Kxb4

64. a4 Kc5 65. a5 Kb4 66. a6 Kc5 67. a7 Kd6 68. a8=R Ke7 69. d6+ Kxd6 70.

e7 Kxe7 71. Kh4 Kd6 72. Kh3 Kc5 73. Kh2 Kd4 74. Kg1 Kxd3 75. Rg8 Ke4 76.

Kh2 Kd5 77. Kh1 Ke5 78. Ra8 Kd5 79. Kg1 Kc6 80. Ra5 Kb6 81. Rh5 Kc7 82. Kh1

Kc6 83. Rh3 Kd5 84. Kg1 Kd4 85. Rh5 Ke3 86. Kh1 Kf3 87. Rb5 Kf4 88. Rb7 Ke4

89. Kg1 Ke5 90. Kf1 Kd4 91. Ke1 Ke5 92. Kd1 Kd5 93. Ra7 Kc5 94. Kc1 Kd5 95.

Kb1 Ke4 96. Ka2 Kd5 97. Ka1 Kd6 98. Ra3 Kd5 99. Kb1 Ke5 100. Kc2 Kd4 101.

Ra5 Kc4 102. Rh5 Kd4 103. Kb3 Ke4 104. Kc3 Kf4 105. Kd4 Kg4 106. Ra5 Kf3

107. Rf5+ Kg4 108. Ke4 Kg3 109. Rf4 Kg2 110. Ke3 Kg3 111. Rd4 Kg2 112. Rg4+

Kh3 113. Kf3 Kh2 114. Kf2 Kh3 115. Ra4 Kh2 116. Rh4# 1-0

[Tadija Mitrovic]

Wow, that Re6 and after that f4 and f5 look like AlphaZero's play. Great!

[plastilino]

How many lost games before this one?

[LuckyDay]

Where did you get that v20 engine from, as only v19 has been released and the devs have not mentioned any release of v20 as far as i know?

The main reason I am asking, is that a few contributors towards training games have been popping up with v20 on their engine listing, which really should not be happening as game contributions should only be done with stable, fully tested engines, and doing so with untested or bootleg engine versions has the risk of sending corrupted training data potentially impacting a training run. Hence i was wondering where these were coming from.

[Frank Müller]

Just one game and SF had to play Black in a crappy Scandinavian??

The meaning of this is guess it : 0.

[Nabil Danial]

I don't know about v.20 but you can always get bleeding edge builds from appveyor. Not recommended of course due to unknown stability status.

[Harbaugh]

In my recent test games, before this one Leela is not losing any, but there were two or three draws.  Leela also lost one on time in a won position! :-(

[Harbaugh]

Hi, The master branch has the v20-dev.  I just issue the following command.

git clone -b master --recurse-submodules https://github.com/LeelaChessZero/lc0.git

[Dietrich Kappe]

Tool for determining Leela Ratio: https://github.com/dkappe/leela-ratio

[bjua...@gmail.com]

@Harbaugh about : "I have both WDL and DTZ installed and Leela detected them both at startup, but it looks like she is only evaluating with the WDL. "

Me too, i'm glad someone else saw the problem. The DTZ files are badly exploited and therefore the game ends are completely polluted.

[Harbaugh]

Just to update, I ran 4 games today (They're about 3 hours each) under net 31461 and they were 4 draws. Openings were English, Sicilian, French.  Leela pressed each time but Stockfish always found a fortress.

[rohit raj]

Thanks for your games. Can you share more games

On Fri 23 Nov, 2018, 8:22 AM Harbaugh <th...@umich.edu wrote:

Just to update, I ran 4 games today (They're about 3 hours each) under net 31461 and they were 4 draws. Openings were English, Sicilian, French.  Leela pressed each time but Stockfish always found a fortress.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/11d723c6-7775-4d44-b5e3-b4a62272de30%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Lothar Jung]

What uci-parameter do you use for multiplexing?

[Harbaugh]

Hi, 

For the backend dropdown I choose multiplexing.

For the BackendOptions textbox I put: a(backend=cudnn-fp16,gpu=0),b(backend=cudnn-fp16,gpu=1)

Here is what the Leela configuration page looks like.

[Dietrich Kappe]

Make sure to share your Leela Ratio. If it’s close to 10, then the network performance is underwhelming.