Leela outplays SF Dev in a position that SF evaluates as draw

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
zullil
Posts: 5452
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: Leela outplays SF Dev in a position that SF evaluates as draw

Post by zullil » Wed Aug 07, 2019 12:35 pm

zullil wrote:
Wed Aug 07, 2019 11:09 am
George Tsavdaris wrote:
Tue Jul 16, 2019 4:28 pm

Lc0 v21.2 JH T8-swa-610000(the TCEC sufi15 winner net):

Code: Select all


 22/55	00:42	 1.448.327	33.759	-0,92	Ra8-d8 Rd1xd8 Rf8xd8 Nb3-d2 Bh5-f7 Re6-e2 Bf7xa2 g2-g3 Bb6-c7 g3xf4 g5xf4 Nf3-h4 Ba2-d5 f2-f3 Kg8-h7 Kg1-f1 Rd8-f8 Bh2-g1 a7-a5 Bg1-c5 Rf8-d8 Nh4-f5 Rd8-d7 Nd2-e4 Bc7-d8 Kf1-g2 Kh7-g6 Nf5-d6 b7-b6 Bc5-f2 Nc6-e5 Bf2-d4 Ne5-c4 Nd6xc4 Bd5xc4 Re2-e1 a5-a4 Ne4-d2 Bc4-f7 Re1-e4 Bd8-c7 Nd2-b1
 22/55	00:43	 1.457.324	33.764	-0,92	Ra8-d8 Rd1xd8 Rf8xd8 Nb3-d2 Bh5-f7 Re6-e2 Bf7xa2 g2-g3 Bb6-c7 g3xf4 g5xf4 Nf3-h4 Kg8-h7 Nh4-g2 Rd8-f8 b2-b3 f4-f3 Nd2xf3 Rf8xf3 Bh2xc7 Ba2xb3 Re2-d2 Rf3xc3 Rd2-d7+ Kh7-g8 Bc7-d6 b7-b5 Rd7-b7 b5-b4 Bd6xb4 Nc6xb4 Rb7xb4 a7-a5 Rb4-b6 a5-a4 Rb6-a6 Rc3-c1+ Kg1-h2 Rc1-a1 Ra6-b6 Ra1-b1 Rb6-a6 h6-h5 Ng2-e3 Rb1-a1 Ra6-b6 Bb3-f7 Rb6-b8+ Kg8-g7 Rb8-b7 a4-a3

Since this thread seems to still be living, here's a Lc0 search of the original position, this time with network 42850 and more than 30,000,000 nodes:

Code: Select all

$ ./lc0
       _
|   _ | |
|_ |_ |_| v0.22.0 built Aug  5 2019
setoption name Backend value cudnn-fp16
setoption name NNCacheSize value 40000000
ucinewgame
Found pb network file: ./network42850
Creating backend [cudnn-fp16]...
GPU: GeForce RTX 2080 Ti
GPU memory: 10.7534 Gb
GPU clock frequency: 1635 MHz
GPU compute capability: 7.5
CUDA Runtime version: 10.1.0
Cudnn version: 7.6.2
Latest version of CUDA supported by the driver: 10.1.0
position fen r4rk1/pp6/1bn1R2p/6pb/5p2/1NP2N1P/PP3PPB/3R2K1 b - - 0 7
go infinite

...

info depth 32 seldepth 73 time 895750 nodes 31677849 score cp 93 hashfull 408 nps 35364 tbhits 0 pv a8d8 d1d8 f8d8 b3d2 h5f7 e6e2 f7a2 g2g3 b6c7 g3f4 g5f4 f3h4 a2d5 f2f3 g8h7 g1f1 d8f8 h2g1 c7d6 e2e1 b7b5 b2b4 d6e5 d2e4 f8d8 e1c1 a7a6 g1b6 d8d7 f1g2 d5e4 f3e4 d7d3 c3c4 d3g3 g2f1 e5b2 c1c2 b2c3 c4b5 a6b5 b6c5 g3h3 h4f5 h7g6 c5d6 c6b4 c2g2 g6f6 d6f4 b4d3
bestmove a8d8 ponder d1d8
quit
This time 50M nodes, but after allowing Lc0 to briefly analyze each prior position in the game:

Code: Select all

HALTED Nodes: 50.2M, N/s: 41.7k, Time: 19m 52s

68.0% Rad8 Rxd8 Rxd8 Nfd4 Bf7 Re1 Nxd4 Nxd4 Bxa2 f3 Bxd4+ cxd4 Rxd4 Kh1 Rd2 b4 Bc4 Bg1 b6 Ra1 Kf8 Bh2 Bb5 Bg1 Kg8 Kh2 Bc6 Kh1 Bb5 Bh2 a5 bxa5 bxa5 Bg1 Rc2 Bd4 a4 Kg1 Kf7 Rb1 Bc6 Rb2 Rc1+ Kh2 h5 Rd2 Kg6 Bb2 Rc4 Rd6+ (N: 49.9M, P: 15.64%)
54.6% Kg7 Rd7+ Rf7 Rdd6 Bc7 Rxh6 Bxd6 Rxh5 Be7 Nxg5 Kg6 Nxf7 Kxh5 Bxf4 Rf8 g4+ Kh4 Ne5 Rxf4 Ng6+ Kg5 Nxf4 Kxf4 Kg2 Ne5 Nd4 Nd3 Ne2+ Ke4 f3+ Ke3 Nd4 Kf4 Ne2+ Ke3 Nd4 Kf4 Ne2+ Ke3 (N: 125k, P: 25.13%)
59.0% Rfd8 Rde1 Bxf3 gxf3 Kg7 h4 Rd5 h5 Rh8 Rg6+ Kf7 Ree6 Ne5 Rgf6+ Kg7 Kg2 Nd7 Rg6+ Kf7 c4 Rd1 c5 Bd8 c6 bxc6 Rxc6 Ne5 Rcd6 Rxd6 Rxd6 Bb6 Nd2 Ke7 Ne4 (N: 64.9k, P: 8.81%)
57.1% Bxf3 gxf3 Rad8 Rxd8 Rxd8 Rxh6 Ne5 Kg2 Kg7 Re6 Ng6 Kf1 Kf7 Re4 Kf6 a4 Rd1+ Re1 Rxe1+ Kxe1 Ne5 Ke2 Nc4 h4 Nxb2 a5 Bc7 hxg5+ Kxg5 Nc5 Kf5 Nxb7 Ke6 Kd2 (N: 22.0k, P: 3.60%)
42.9% Rae8 Rxh6 Bxf3 gxf3 Kg7 Rhd6 Ne5 Kg2 Ng6 Kf1 Ne5 Kg2 Ng6 Rd7+ Rf7 Kf1 Ne5 Rxf7+ Kxf7 Nd2 Rh8 Ne4 Nxf3 Kg2 Nh4+ Kf1 Nf3 (N: 16.3k, P: 6.16%)
32.7% Kh7 Rdd6 Kg7 Rxh6 Bxf3 gxf3 Ne5 Kg2 Rad8 h4 Rxd6 Rxd6 gxh4 Nd2 Bc7 Rd4 Nc6 Rd7+ Rf7 Rxf7+ Kxf7 Kh3 Ke6 Kg4 Ne5+ Kxh4 Kf5 Kh3 Nd3 (N: 7.8k, P: 4.16%)
38.1% Rfe8 Rxh6 Rad8 Rxd8 Rxd8 g4 Bf7 Nbd4 Nxd4 cxd4 Rd5 h4 gxh4 Bxf4 Bxd4 Nxd4 Rxd4 Rf6 Bxa2 Kh2 Rb4 Kh3 Rxb2 Be3 a5 Kxh4 a4 g5 a3 Bc5 Rb3 (N: 7.6k, P: 3.45%)
26.2% Bf7 Rxh6 Rad8 Rhd6 Rxd6 Rxd6 Re8 g4 Re2 Rd2 Bxf2+ Kf1 Rxd2 Nbxd2 Be3 b3 Bxd2 Nxd2 Kg7 Ne4 Kg6 Bg1 b6 c4 Nb4 (N: 4.9k, P: 3.09%)
42.8% Bc7 Rxh6 Rad8 Rxd8 Rxd8 g4 Bf7 Nxg5 Rd1+ Kg2 Bd5+ f3 Rb1 Rg6+ Kh8 Bg1 Rxb2+ Kh1 Rb1 c4 Bg8 h4 Bb6 c5 Bc7 (N: 4.0k, P: 1.52%)
39.3% Kf7 Rxh6 Bxf3 gxf3 Ne5 Kg2 Rh8 Rhd6 Rad8 Rxd8 Rxd8 Rxd8 Bxd8 h4 Nd3 hxg5 Bxg5 Na5 b6 Nc6 a5 b3 Ke6 (N: 3.7k, P: 1.62%)

jp
Posts: 651
Joined: Mon Apr 23, 2018 5:54 am

Re: Leela outplays SF Dev in a position that SF evaluates as draw

Post by jp » Thu Aug 08, 2019 4:12 am

George's Lc0 v21.2 JH T8-swa-610000 (the TCEC sufi15 winner net) output:



Louis's Lc0 v0.22.0 42850 output (50M nodes, after allowing Lc0 to briefly analyze each prior position in the game):




In the Leela NN 42850 PV after 13... Rxd4, SF's eval is -1.09 (depth=37), but it's surely a draw. Its line ends in the position


jp
Posts: 651
Joined: Mon Apr 23, 2018 5:54 am

Re: Leela outplays SF Dev in a position that SF evaluates as draw

Post by jp » Thu Aug 08, 2019 7:38 am

Lc0 v0.22.0 42850's 2nd & 3rd lines:




jp
Posts: 651
Joined: Mon Apr 23, 2018 5:54 am

Re: Leela outplays SF Dev in a position that SF evaluates as draw

Post by jp » Fri Aug 16, 2019 7:35 am

In Lc0 v0.22.0 42850's 2nd PV after 9. gxf3 Kg7,



SF's eval is (depth=48)
0.00 10.a4 Rd5 11.h4 Rad8 12.Kg2 R8d6 13.R6e2 Rd3 14.Re8 a5 15.R1e4 R3d5 16.h5 Rd8 17.R8e6 R5d6 18.Re2 Kf7 19.R6e4 R8d7 20.Bg1 Rd1 21.Re6 R1d6

After 24. Ne4,



SF's eval is (depth=50)
0.00 24...Bc7 25.Rf6 Nd3 26.Kf1 Ne5

In Lc0 v0.22.0 42850's 3rd PV after 10. Rxh6,



SF's eval is (depth=50)
0.00 10...Ne5 11.Kg2 Kg7 12.Re6 Ng6 13.Kf1 Kf7 14.Re4 Rd1+ 15.Re1 Rd8

After 24. Kd2,



SF's eval is (depth=60)
0.00 24...Kd5 25.Kc2 Na4 26.Kb3 Nc5+ 27.Nxc5 Kxc5 28.a6 Kb5 29.c4+ Kxa6 30.Kc3 Kb6 31.Kb4 Bd6+ 32.Ka4 Kc5 33.Ka5 Kxc4 34.Ka6 Be5 35.Bxf4 Bxf4 36.Kxa7 Kb5 37.Kb7 Kc5 38.Ka6 Kc6 39.Ka5 Be5 40.Ka6 Bc7 41.f4 Bxf4 42.Ka5 Bd2+ 43.Ka6 Kc5 44.Kb7 Bf4 45.f3 Kb5 46.Ka8 Bd6 47.Ka7 Kc6 48.Ka6 Bc7 49.f4 Bxf4 etc.


So in all of Leela's top 3 lines, it looks like a draw.

Everywhere we look, it looks like a draw. If there's a win for Black, SF & Lc0 cannot find it.

Post Reply