PHCEL and LC0

[Rubinus]

LCZero gauntlets 2+2
12x (Fritz 16, Nimzo 8, Senpai 2, Colosss 2008b, Hiarcs 14, ProDeo 2.6, Spike 1.4, Fruit 2.3.1, Junior Yokohama, Zappa II Mexico)
id 359 - 51.0/120, ph-elo 2943 (120g)
id 390 - 48.5/120, ph-elo 2928 (120g)
games 359, games 390
Next id 483

All my LC0 games

id 483 - 49.0/120, ph-elo 2931 (120g), games 483
Other
LCzero v0.10-id338 ph-elo 2867 (20g - swiss A 004/2018)
LCzero v0.8-id253 ph-elo 2815 (39g - swiss B+A 004/2018)

[Rubinus]

LCZero 0.10-id484 - Spike 1.4_SpikeC80-C83Book_12m+5s +4=7-7
games
There were two extra games in error and LCZero both lost ... It seems that the loss after restart is deleted, id 484 and even better 485 are probably better than id 359 before restart. Even Moses wandered for 40 days of desert ... Ted just when we came back to the ceiling? I'm afraid that without the program changes, it will not replace tactical blindness.SpikeC80-C83Book
Next gauntlet id 485, match from C84-C89.

All my LC0 games

[Rubinus]

LCZero 0.10-id494 - Spike 1.4_SpikeC80-C83Book_12m+5s +4=7-7
games
There were two extra games in error and LCZero both lost ... It seems that the loss after restart is deleted, id 494 and even better 495 are probably better than id 359 before restart. Even Moses wandered for 40 days of desert ... Ted just when we came back to the ceiling? I'm afraid that without the program changes, it will not replace tactical blindness.SpikeC80-C83Book
Next gauntlet id 495, match from C84-C89.

All my LC0 games

[Rubinus]

LCZero gauntlets 2+2
12x (Fritz 16, Nimzo 8, Senpai 2, Colosss 2008b, Hiarcs 14, ProDeo 2.6, Spike 1.4, Fruit 2.3.1, Junior Yokohama, Zappa II Mexico)
id 359 - 51.0/120, ph-elo 2943 (120g)
id 483 - 49.0/120, ph-elo 2931 (120g),
id 390 - 48.5/120, ph-elo 2928 (120g)
other:
LCzero v0.10-id338 ph-elo 2867 (20g - swiss A 004/2018)
LCzero v0.8-id253 ph-elo 2815 (39g - swiss B+A 004/2018)


games 359, games 390, ), games 483

All my LC0 games

id 495 - 58.5/120, ph-elo 2987 (120g), games 495

[Rubinus]

LCZero 0.10-id501 - Spike 1.4_SpikeC84-C89Book_12m+5s +8=4-12
games
Because there was no time to update the new version for Swiss 5, there have been some matches. Network 501 is good, but if it exceeds previously used 495, it's hard to guess.
SpikeC84-C89Book
Next match from C90-C95.

LCZero 0.10-id501 - Arasan 20.5_12m+5s +0=1-5
games
Arasan use own binary book. Heavy debacle, but few games, maybe LCZ caught time. Just filling in the time, there was no reason to continue with Arasan's version of 21.0.

LCZero 0.10-id501 - Scorpio 2.8.6_12m+5s +8=2-6
games
Both with my test book, so the variants were always played on both sides. Scorpio also has some knowledge network so I left MCS and net6 on. She chose the fifth option when LCZ won both sides, otherwise identical results.

All my LC0 games

[Rubinus]

Ouu!
[pgn][Event "Bswiss_2+2, ph-0005-2018"] [Site "Prague"] [Date "2018.07.28"] [Round "6"] [White "Petir 4.9999"] [Black "LCzero v0.10-id501"] [Result "1-0"] [ECO "B07"] [PlyCount "35"] [EventDate "2018.07.28"] 1. e4 d6 2. d4 Nf6 3. f3 c5 4. c3 e6 5. dxc5 d5 6. b4 a5 7. Nh3 axb4 8. cxb4 b6 9. exd5 Nxd5 10. Nf4 Nxb4 11. cxb6 Qf6 12. Bb5+ Bd7 13. b7 Ra7 14. Qxd7+ Nxd7 15. b8=Q+ Ke7 16. Qxa7 Qe5+ 17. Be3 Kf6 18. Bxd7 1-0[/pgn]

[Rubinus]

LCZero 0.10-id501 - Spike 1.4_SpikeC90-C95Book_12m+5s +5=2-9
games
SpikeC90-C95Book
Next two versions tournament book.

All my LC0 games

[Rubinus]

Spike against Spike with the SpikeTurnPH-Book library (the tournament book I developed in 2006-8, I quite won with it), 12m + 5s. Just three matches in one. LC0 came with program update 0.16. So I installed the opencl because I have a Radeon card and not NVidia. It goes with the client, but when it runs in the GUI, it starts to produce error messages after a while.
Spike 1.4 - LCZero 0.16opencl id 532 1.5: 0.5 + 1 / = 1 / -0 The second bout lost to time, the third LC0 did not start the game, it ended.
Spike 1.4 - LCzero v0.10-id532 16.5: 15.5 + 11 / = 11 / -10 here the network has confirmed the quality and with the old version of the program 0.10 against the best book the best result.
Spike 1.4 - LCZero 0.16blas id 532 14.0: 2.0 + 12 / = 4 / -0, so another compilation of 0.16. It works, but the result is clear trouble. Four times lost, one in a winning position. Slightly depth. Apparently something was wrong. Syzygy and VisibleThinking support disappeared. I'm upset. I still have to use 0.10.

games
SpikeTurnPH-Book (last version, 2007 or 8)

All my LC0 games

LC0 0.16 - Without analyze, syzygy, time limit problem, shit compilation.

[Rubinus]

LCZero 0.10-id570 - Chiron 4 64bit_12m+5s +0=4-8 Quite a bit, but LC0 with the last nets begins to match Spike, Chiron is 200 ELO higher.
Chiron use own binary book.
games

All my LC0 games

[Rubinus]

Chiron 4 64bit - Lc0 v0.17.0-rc2-blas-id10968_12m+5s +9=9-2
games
Chiron use own binary book.
That's better. It took me for a while to find a new version of the new version that would not work, I currently have a version of opencl but the setting must be NN backend to use = 1, so blas. At 0, opencl, my engine falls ... Another change against the standard is Minibatch size for NN inference = 16, the original value 256 exceeds the time.