Page 2 of 3
Re: Leela on a weak pc, question
Posted: Sat Jan 12, 2019 2:20 am
by Eduard
Thank You!!
Re: Leela on a weak pc, question
Posted: Sat Jan 12, 2019 9:29 am
by Tibono
dkappe wrote: ↑Sat Jan 12, 2019 1:54 am
64x6 or 64x8?
64x8
. No 64x6 hosted behind the link I provided.
dkappe wrote: ↑Sat Jan 12, 2019 1:54 am
The 96x8-se is already released.
Thanks! Looking forward to test it!
Re: Leela on a weak pc, question
Posted: Sat Jan 12, 2019 12:38 pm
by Eduard
Asus Laptop 2x2,4 GHz 64 Bit, my first tests:
Lc0 v20.1 for windows only, with
Network 11258-16x2-se-4.pb, fastest network.
Network 11258-96x8-se-5.pb.gz, is little bit slower than 16x2, but all in all clear much stronger! Best network for my PC.
Network 128x10 is currently to slow for my PC. I works, but is to slow.
Newertheless I let play some games with the network 16x2. Lc0 played vs. old star Chess Genius runs on Smartphone (400 Kn/s).
On timecontrol 30m/game and ponder ON, Lc0 (with 16x2) won clear 2-0.
Here two interesting diagrams
Game 1: Chess Genius - Lc0 v20.1 11258-16x2
[d]5r2/2k1r2p/3p2nb/2pP4/ppP1p3/PP3pP1/2N2P1P/1N2RRK1 w - - 0 30
It was intersesting to see how Lc0 pushes its pawns forward. Chess Genius had no chance.
Game 2: Lc0 v20.1 11258-16x2 - Chess Genius
[d]1r3rk1/2nnbpp1/1q1p4/p1pPpPN1/2P1P2P/6R1/1PB1Q1K1/2B4R b - - 0 30
Chess Genius sees no danger. That is the point here! The hardware of Genius was not bad, really. But the engine was blind.
Lc0 attacks and sees itself clearly in the advantage +4!
Fallowing blunder of Lc0 at the TCEC against Stockfish, I can not reproduce with Lc0 v20.1 11258-96x8-se.
Lc0 at TCEC played 33...Qf4?
[d]6k1/2q2pp1/p4n1p/1prp4/2nN1P1P/3B1P2/PP3QP1/3R2K1 b - - 0 1
Analysis by Lc0 v0.20.1 11258:
33...Qxf4 34.Ne6 fxe6 35.Qxc5 Nxb2 36.Bg6 Nd7 37.Qd4 Qxd4+ 38.Rxd4 Ne5 39.h5 Kf8 40.f4 Nxg6 41.hxg6 Nc4 42.f5 Ke7 43.g4 Kf6 44.Kf2 Ke5
= (0.22) Tiefe: 9/25 00:01:33 18kN
33...Nb6 34.Nb3 Rc6 35.g3 Na4 36.Kg2 Nc5 37.Nxc5 Rxc5 38.Re1 Rc1 39.Rxc1 Qxc1 40.Qc2 Qxc2+ 41.Bxc2
= (0.16) Tiefe: 9/25 00:01:34 19kN
33...Nb6 34.Nb3 Rc6 35.g3 Na4 36.Kg2 Nc5 37.Nxc5 Rxc5 38.Qd4 h5 39.Re1 Rc6 40.Qe5 Qc8 41.Bf5 Qd8
= (0.16) Tiefe: 9/25 00:01:39 20kN
Re: Leela on a weak pc, question
Posted: Sun Jan 13, 2019 2:39 am
by Eduard
First a small correction: Network "128x10-se-distill-ccrl-11248.pb.gz" is not slow on my PC. It works fine. Sorry.
In my first games, Lc0 losed 3 games unfortunately in the endgame, after good middlegame advantage. Lc0 played fantastic, but then came the endgame. For most other Engines following endings are not to hard or difficult.
Lc0 v20.1 played with network 11258-96x8-se-5.pb.gz:
Position/Game 1
Lc0 v0.20.1 - Rybka 4.1 x64 0-1, ASUS-PC, Schnellschach 20m+5s 2019
[d]3k4/1R6/7p/p4P2/4Pr2/1P2K2p/8/8 w - - 0 1
Lc0 played now 55. Kxf4?? and losed very soon!
Analysis by Lc0 v0.20.1:
55.Kxf4 h2 56.f6 h1Q 57.f7 Qc1+ 58.Ke5 Qc3+ 59.Ke6 Qh3+ 60.Kf6 Qh4+ 61.Kg7 Qg5+ 62.Kh7 Qf6 63.Rb8+ Kc7 64.f8Q Qxf8 65.Rxf8
=/+ (-0.42) Tiefe: 13/27 00:01:08 29kN
55.Kxf4 h2 56.f6 h1Q 57.f7 Qc1+ 58.Ke5 Qc3+ 59.Ke6 Qh3+ 60.Kf6 Qh4+ 61.Kg7 Qg5+ 62.Kh7 Qf6 63.Rb8+ Kc7 64.f8Q Qxf8 65.Rxf8
=/+ (-0.45) Tiefe: 13/27 00:01:13 31kN
55.Kxf4 h2 56.f6 h1Q 57.f7 Qc1+ 58.Ke5 Qc3+ 59.Ke6 Qh3+ 60.Kf6 Qh4+ 61.Kg7 Qg5+ 62.Kh7 Qf6 63.Rb8+ Kc7 64.f8Q Qxf8 65.Rxf8
=/+ (-0.45) Tiefe: 13/27 00:01:18 35kN
Position/Game2:
Rybka 4.1 x64 - Lc0 v0.20.1 1-0, ASUS-PC, Schnellschach 20m+5s 2019
[d]8/6p1/4Qpk1/p4n1p/2P2P2/1r1R2PP/2q1N1K1/8 w - - 0 1
Now is Rybka to move. Last move of Lc0 was 53...Qc2??
Analysis by Lc0 v0.20.1:
54.Qe8+ Kh7 55.Rxb3 Qxb3 56.Qxh5+ Nh6 57.Qxa5 Qxc4 58.Kf2 Nf7 59.Qc3 Qd5 60.Qf3 Qc5+ 61.Kg2 Nd6 62.Qd3+ f5 63.g4 Qc6+
+/- (1.03) Tiefe: 13/22 00:01:33 35kN
54.Qe8+ Kh7 55.Rxb3 Qxb3 56.Qxh5+ Nh6 57.Qxa5 Qxc4 58.Kf2 Nf7 59.Qc3 Qd5 60.Qf3 Qc5+ 61.Kg2 Nd6 62.Qd3+ f5 63.g4 Qc6+
+/- (1.03) Tiefe: 14/22 00:01:34 36kN
Lc0 missed here the following move. Because, Rybka did not play 54. Qe8++?! but 54. Rxb3!! with fast win in the endgame.
Is there a way to better train Lc0 for such positions with such small networks?
Re: Leela on a weak pc, question
Posted: Sun Jan 13, 2019 3:19 pm
by dkappe
If you want a small net that plays a better endgame, try Ender.
https://github.com/dkappe/leela-chess-w ... ndgame-Net
I’m going to distill it down from 128x10 to 112x9 when my big distillation is done.
It trained with egtb, so it’s designed to play 16p or less with 6p tb.
Re: Leela on a weak pc, question
Posted: Sun Jan 13, 2019 3:21 pm
by dkappe
Note, Ender only plays endgames. It’s opening and middlegame play is laughable.
Re: Leela on a weak pc, question
Posted: Sun Jan 13, 2019 9:08 pm
by Eduard
Thank You!
Yes that's it.
Is it possible to nun networks in parallel? Or, that the engine automatically selects another network at a certain number of moves?
Lc0 v0.20.1 - Rybka 4.1 x64 0-1, ASUS-PC, Schnellschach 20m+5s 2019
3k4/1R6/7p/p4P2/4Pr2/1P2K2p/8/8 w - - 0 1
Analysis by Lc0 v0.20.1 (ender):
55.Rb6 Rh4 56.Rd6+ Ke7 57.Rd1 h2 58.Rh1 Kf6 59.Kf3 Kg5 60.Kg3 Rxe4 61.Rxh2 Rb4 62.f6 Rg4+ 63.Kf3 Rf4+
=/+ (-0.63) Tiefe: 10/19 00:00:30 3kN
Re: Leela on a weak pc, question
Posted: Sun Jan 13, 2019 9:43 pm
by dkappe
Someone made a patch for lc0 v19.1 that switched nets at a certain piece count and changed cpuct, policy softmax and fpur at the same time. I don't think it's been ported to v20.x
Also, just released 11258-112x9-se. Best one yet.
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%)
1 ethereal : 3360 51 290.5 320 90.8 100 267 47 6 14.7
2 ID11258-112x9-se : 3098 58 67.0 104 64.4 91 49 36 19 34.6
3 crafty25.2 : 3051 32 230.5 363 63.5 69 189 83 91 22.9
4 ID36092 : 3036 49 88.0 146 60.3 59 58 60 28 41.1
5 ID11258-96x8-se-5 : 3029 34 169.0 296 57.1 55 124 90 82 30.4
6 ID35975 : 3026 39 145.5 252 57.7 66 103 85 64 33.7
7 ID11248-128x10-se : 3016 31 220.5 380 58.0 79 167 107 106 28.2
8 ID35689 : 2995 44 103.5 175 59.1 52 70 67 38 38.3
9 cheng4 : 2993 65 37.0 73 50.7 85 27 20 26 27.4
10 amoeba : 2949 50 68.0 144 47.2 61 45 46 53 31.9
11 scs-64x8-run1-550000 : 2938 60 42.5 96 44.3 100 31 23 42 24.0
12 ID11248-256x12-se : 2808 54 41.0 126 32.5 62 22 38 66 30.2
13 winter : 2796 54 34.5 129 26.7 100 21 27 81 20.9
14 crafty19.18 : 2693 38 79.0 363 21.8 66 53 52 258 14.3
15 ID11258 : 2679 64 26.0 124 21.0 95 15 22 87 17.7
16 ID11258-16x2-se-3 : 2592 73 21.5 126 17.1 58 11 21 94 16.7
17 ID11258-16x2-se-4 : 2583 66 24.0 159 15.1 --- 12 24 123 15.1
White advantage = 45.21 +/- 8.08
Draw rate (equal opponents) = 36.60 % +/- 1.51
Re: Leela on a weak pc, question
Posted: Sun Jan 13, 2019 10:19 pm
by Eduard
dkappe wrote: ↑Sun Jan 13, 2019 9:43 pm
Also, just released 11258-112x9-se. Best one yet.
Very nice! The first analyzes shows all in all best results of these Networks. Lc0 is also good fast on my PC with this Network. Thank You!
Re: Leela on a weak pc, question
Posted: Mon Jan 14, 2019 12:06 am
by Eduard
Here is the first game with this new network "11258-112x9-se.pb.gz."
I saw a fantastic game of Lc0. Game was over after only 29 moves!
Both engines played with a short 5 move book, and with 1 Core 2,4 Ghz.
Rybka is only single core engine, and so I let Lc0 run on 1 Core.
[Event "ASUS-PC, Schnellschach 20m+5s"]
[Site "ASUS-PC"]
[Date "2019.01.13"]
[Round "1"]
[White "Lc0 v0.20.1"]
[Black "Rybka 4.1 x64"]
[Result "1-0"]
[ECO "D43"]
[Annotator "0.26;-0.13"]
[PlyCount "57"]
[TimeControl "1200+5"]
{Intel(R) Core(TM) i3 CPU M 370 @ 2.40GHz 2466 MHz W=9.5 plies; 0kN/s;
engPowerbook.ctg B=15.3 plies; 100kN/s; Eduard.ctg} 1. c4 {B 0} Nf6 {B 0} 2. d4
{B 0} e6 {B 0} 3. Nc3 {B 0} d5 {B 0} 4. Nf3 {B 0} c6 {B 0} 5. Bg5 {B 0} h6 {
B 0 Beide letzter Buchzug} 6. Bh4 {0.26/7 21} dxc4 {-0.13/16 49} 7. e4 {
0.25/9 19 (a4)} g5 {-0.25/17 45} 8. Bg3 {0.29/9 19} b5 {-0.25/16 52} 9. h4 {
0.28/8 41 (a3)} g4 {-0.39/15 39} 10. Ne5 {0.22/9 19} h5 {-0.39/15 45} 11. Be2 {
0.28/8 27 (a3)} b4 {-0.34/13 33} 12. Na4 {0.38/9 27} Nxe4 {-0.34/15 46} 13. O-O
{0.51/9 25} Bg7 {-0.34/15 45} 14. Bf4 {0.55/8 28} Qxh4 {-0.34/14 54 (Nd6)} 15.
Bxc4 {0.92/9 40} g3 {0.00/14 29 (Qf6)} 16. fxg3 {1.73/10 38} Nxg3 {0.29/15 22}
17. Qf3 {1.81/9 22 (Bxg3)} Qh1+ {0.00/15 57 (Nf5)} 18. Kf2 {1.72/13 0} Nxf1 {
0.00/17 36} 19. Rxf1 {2.32/10 46} Qh4+ {0.00/16 37} 20. Kg1 {2.07/12 32} O-O {
0.00/17 42 (Nd7)} 21. Qe3 {2.61/9 44} Bxe5 {0.00/16 35} 22. dxe5 {2.25/10 51
(Bxe5)} Ba6 {0.24/13 73} 23. Rf3 {2.90/13 16} Bxc4 {2.31/13 19 (Rd8)} 24. Bg5 {
6.08/9 49} Qg4 {2.31/14 31} 25. Rg3 {6.97/8 21} f5 {2.31/13 30 (Qd1+)} 26. Rxg4
{8.29/9 65 (Bf6)} hxg4 {10.19/15 71} 27. Bf6 {9.19/8 9} Kf7 {10.19/15 27} 28.
Nc5 {10.34/7 29 (Qh6)} Rh8 {7.33/14 21 (Ke8)} 29. Bxh8 {16.71/6 33} 1-0
A real good game of Lc0 only running on CPU.