Closed MelleKoning closed 4 years ago
Another example from another lost game from the same test from Ingo: See move 118.Rf2 in this game, which looks like a total draw for over 50 moves before until leela decides to go into a wrong direction on move 118: https://lichess.org/lZPrcLkU#235
Another net from test40, 40110 also gives a blunder move. This time not after 10k but after 20k nodes.
go nodes 20000
info depth 6 seldepth 6 time 2 nodes 8837 score cp 207 hashfull 31 nps 500 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 4 seldepth 6 time 16 nodes 8838 score cp 207 hashfull 31 nps 125 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 5 seldepth 7 time 83 nodes 8839 score cp 207 hashfull 31 nps 36 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 4 seldepth 7 time 96 nodes 8841 score cp 207 hashfull 31 nps 52 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 5 seldepth 7 time 107 nodes 8844 score cp 207 hashfull 31 nps 74 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 5 seldepth 9 time 121 nodes 8850 score cp 207 hashfull 31 nps 115 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 6 seldepth 9 time 287 nodes 8874 score cp 207 hashfull 31 nps 132 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 6 seldepth 10 time 532 nodes 8923 score cp 207 hashfull 31 nps 163 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 6 seldepth 11 time 540 nodes 8927 score cp 207 hashfull 31 nps 168 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 6 seldepth 12 time 593 nodes 8928 score cp 207 hashfull 31 nps 155 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 6 seldepth 13 time 646 nodes 8935 score cp 207 hashfull 31 nps 153 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 7 seldepth 13 time 654 nodes 8948 score cp 207 hashfull 31 nps 171 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 7 seldepth 14 time 1024 nodes 9001 score cp 207 hashfull 32 nps 161 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 7 seldepth 15 time 2240 nodes 9149 score cp 207 hashfull 32 nps 139 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 7 seldepth 15 time 7264 nodes 10385 score cp 203 hashfull 35 nps 213 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5 c6c8 b5b6
info depth 7 seldepth 16 time 8080 nodes 10530 score cp 205 hashfull 36 nps 209 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 7 seldepth 17 time 9340 nodes 10809 score cp 206 hashfull 36 nps 211 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 8 seldepth 17 time 9488 nodes 10865 score cp 206 hashfull 36 nps 213 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 8 seldepth 17 time 14512 nodes 11567 score cp 205 hashfull 40 nps 188 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 8 seldepth 17 time 19514 nodes 12473 score cp 204 hashfull 43 nps 186 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 7 seldepth 17 time 23696 nodes 13307 score cp 201 hashfull 45 nps 188 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 7 seldepth 17 time 28725 nodes 14661 score cp 201 hashfull 48 nps 202 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 7 seldepth 17 time 33752 nodes 16105 score cp 201 hashfull 50 nps 215 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 7 seldepth 17 time 38767 nodes 17174 score cp 201 hashfull 52 nps 215 tbhits 0 pv h1d5 g8g7 f4c1 g7h7 c1f4 h7g7 f4e3 g7h7 e3c1 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5
info depth 7 seldepth 17 time 39434 nodes 17284 score cp 220 hashfull 52 nps 214 tbhits 0 pv f4c1 g8g7 h1d5 a6c6 c1f4 g7h7 d4e3 h7g7 e3d2 g7h7 d2e2 h7g7
info depth 7 seldepth 17 time 44462 nodes 17840 score cp 217 hashfull 52 nps 202 tbhits 0 pv f4c1 g8h7 h1d5 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5 g7h7 d4e3 h7g7 e3f2 g7h7 f2f3
info depth 7 seldepth 17 time 49472 nodes 18326 score cp 216 hashfull 53 nps 191 tbhits 0 pv f4c1 g8h7 h1d5 h7g7 c1f4 g7h7 d5f3 h7g7 f3d5 g7h7 d4e3 h7g7 e3f2 g7h7 f2f3
info depth 7 seldepth 17 time 51919 nodes 18644 score cp 216 hashfull 53 nps 188 tbhits 0 pv f4c1 g8h7 h1d5 h7g7 c1d2 g7h7 d2f4 h7g7 c5c6 a6c6 d5b5 g7f7 b5d5 f7g7 b4b5
info string h1a1 (171 ) N: 6 (+ 0) (P: 1.67%) (Q: -0.89682) (U: 1.41453) (Q+U: 0.51771) (V: -0.4852)
info string f4e5 (829 ) N: 7 (+ 0) (P: 0.83%) (Q: -0.14328) (U: 0.61293) (Q+U: 0.46964) (V: -0.3229)
info string f4g5 (831 ) N: 8 (+ 0) (P: 0.79%) (Q: -0.07812) (U: 0.51846) (Q+U: 0.44034) (V: -0.0435)
info string h1a8 (192 ) N: 9 (+ 0) (P: 2.19%) (Q: -0.82761) (U: 1.29703) (Q+U: 0.46942) (V: 0.2662)
info string h1c6 (188 ) N: 9 (+ 0) (P: 2.34%) (Q: -0.86421) (U: 1.38425) (Q+U: 0.52004) (V: -0.1015)
info string h1e4 (184 ) N: 10 (+ 0) (P: 2.61%) (Q: -0.90933) (U: 1.40673) (Q+U: 0.49740) (V: -0.3522)
info string h1b7 (190 ) N: 10 (+ 0) (P: 2.52%) (Q: -0.82284) (U: 1.35907) (Q+U: 0.53623) (V: 0.3690)
info string g3g4 (615 ) N: 18 (+ 0) (P: 1.08%) (Q: 0.09576) (U: 0.33704) (Q+U: 0.43280) (V: 0.1677)
info string h1h3 (183 ) N: 22 (+ 0) (P: 1.39%) (Q: 0.10933) (U: 0.35781) (Q+U: 0.46714) (V: -0.1393)
info string d4d5 (761 ) N: 43 (+ 0) (P: 8.00%) (Q: -0.54412) (U: 1.07682) (Q+U: 0.53270) (V: 0.1988)
info string h1h2 (180 ) N: 59 (+ 0) (P: 2.47%) (Q: 0.20118) (U: 0.24381) (Q+U: 0.44499) (V: -0.0315)
info string h1c1 (173 ) N: 61 (+ 0) (P: 4.43%) (Q: 0.05528) (U: 0.42350) (Q+U: 0.47878) (V: 0.0142)
info string h1f1 (176 ) N: 62 (+ 0) (P: 2.99%) (Q: 0.17796) (U: 0.28084) (Q+U: 0.45880) (V: 0.0650)
info string h1g1 (177 ) N: 62 (+ 0) (P: 3.83%) (Q: 0.10973) (U: 0.35968) (Q+U: 0.46941) (V: 0.0440)
info string h1b1 (172 ) N: 62 (+ 0) (P: 3.92%) (Q: 0.10389) (U: 0.36814) (Q+U: 0.47203) (V: 0.0767)
info string h1d1 (174 ) N: 76 (+ 0) (P: 4.84%) (Q: 0.09964) (U: 0.37211) (Q+U: 0.47175) (V: 0.1567)
info string c5c6 (973 ) N: 85 (+ 0) (P: 2.87%) (Q: 0.24485) (U: 0.19748) (Q+U: 0.44233) (V: 0.3303)
info string d4e3 (750 ) N: 179 (+ 0) (P: 3.11%) (Q: 0.32429) (U: 0.10249) (Q+U: 0.42677) (V: 0.2531)
info string h1e1 (175 ) N: 526 (+ 0) (P: 9.16%) (Q: 0.32405) (U: 0.10301) (Q+U: 0.42706) (V: 0.5120)
info string f4e3 (817 ) N: 864 (+ 0) (P: 2.67%) (Q: 0.39004) (U: 0.01828) (Q+U: 0.40832) (V: 0.2430)
info string h1g2 (179 ) N: 1267 (+ 0) (P: 8.11%) (Q: 0.37566) (U: 0.03788) (Q+U: 0.41354) (V: 0.3393)
info string f4h6 (837 ) N: 1623 (+ 0) (P: 2.38%) (Q: 0.39871) (U: 0.00870) (Q+U: 0.40741) (V: 0.3415)
info string h1f3 (181 ) N: 1771 (+ 0) (P: 8.86%) (Q: 0.38212) (U: 0.02963) (Q+U: 0.41175) (V: 0.4051)
info string f4d2 (811 ) N: 1838 (+ 0) (P: 1.74%) (Q: 0.40046) (U: 0.00559) (Q+U: 0.40605) (V: 0.2403)
info string h1d5 (186 ) N: 4303 (+ 0) (P: 13.04%) (Q: 0.39129) (U: 0.01794) (Q+U: 0.40923) (V: 0.5185)
info string f4c1 (809 ) N: 5663 (+ 0) (P: 2.18%) (Q: 0.41320) (U: 0.00228) (Q+U: 0.41548) (V: 0.2134)
bestmove f4c1 ponder g8h7
The move Bf4-c1 is also a blunder move while Qh1-d5 which is found before is the move that would hold the draw. Suspect therefore that this is not a net issue, but some search- or cutoff issue for game-end although I would not know how to prove that.
Try this version with Certainty Propagation enabled: https://ci.appveyor.com/project/LeelaChessZero/lc0/builds/21908253/job/92pkh2tvbq9ffihg/artifacts
Or the stronger version with look-ahead here, but you might need to increase threads to ( 2 * no. of gpus) +1: https://ci.appveyor.com/project/LeelaChessZero/lc0/builds/21916156/job/851awkkkw5etj92n/artifacts
both are the cuda versions, but others are available too. These versions might help in some of your positions.
Occasional blind spots and blunders are to be expected from the nature of search and training; the frequency has dropped significantly though, so this isn't a real issue anymore. Issue can be closed.
Closing issue
Found in the testgames of Ingo the following loss from network 32367 against stockfish10.
What surprises me most is the fact that the game is stalled for a very long time, both sides just shuffling their pieces around while nothing changes in the pawn-structure, until Lc0 suddenly chooses a blunder-move that loses the game.
Q: How come this blunder is suddenly occuring?
Game can be found on Lichess: https://lichess.org/3M9fwD1H#164
The blunder move played is 83.Bf4-h6, while lots of other moves would keep the draw (as all the earlier moves played already) How come move f4h6 is preferred? Does leela not see the counter-move from stockfish Rxd6+!!, while it apparently was aware of that counter move all the previous moves all along?
To reproduce: