Allie's Missed wins TCEC S18

[jjoshua2]

There's 3-5 missed wins now, with most people saying 4. I've analyzed a lot of positions to 60+ and 70+ depth.

Aloril analyzed the recent Fire - Allie game and got exf3 down to a tb win. I analyzed Leelenstein same move to 20 million nodes with two different sets of search parameters and got it wanting Nb6 both times, but similar eval to Ke7 and above exf3 which was still close. Allie played Ke7 in game which draws.

Code: Select all

Fire 8_beta - AllieStein v0.7_dev-net_14.3 Line, TCEC Season 18 - Division P 2020
1b6/1P4p1/1nP2k2/1P3p2/2B1pP2/6P1/6K1/8 w - - 0 1

Analysis by Brainfish 280520 64 POPCNT:

48.Bb3 Ke7 49.Kf2 Kd6 50.Bf7 Nd5 51.Bg6 Ke6 52.Bh5 Ke7 53.Bg6 Kf6 54.Bh7 g5 55.Bg8 Ne7 56.fxg5+ Kxg5 57.Bf7 Kf6 58.Be8 Nd5 59.Bd7 Bd6 60.Kg2 Bc7 61.Be8 Ke5 62.Bd7 Kd6 63.Kf2 Ne7 64.Ke2 Kc5 65.g4 f4 66.g5 Kb6 67.Ke1 Bb8 68.Kf1 Be5 69.Ke2 Kc7 70.Bh3 Bd6 71.Bg4 Bc5 72.Bh5 Bb6 73.Ke1 f3 74.g6 Bd4 75.Kd1 Bc5 76.g7 Bd4 77.Bg6 e3 78.Be4 e2+ 79.Kd2 Bc3+ 80.Kc2 Ba5 81.b6+ Bxb6 
  Black is clearly winning: -+ (-9.33 --)  Depth: 56/79   03:18:03  164681MN, tb=387821157

Looks like SF agrees that Nb6 also would have won.

[jjoshua2]

Allie - SF seems to have probably 3 more missed wins this game it didn't play. with eval slowly dropping.

Code: Select all

AllieStein v0.7_dev-net_14.3 - Stockfish 202005232210 1/2-1/2, TCEC Season 18 - Division P 2020
Analysis by Brainfish 280520 64 POPCNT:
53.Bf1 Kg7 54.Qe2 Kh8 55.Bg2 Kg7 56.Kb2 Ne7 57.Rh1 Qb7 58.Rf1 Ng8 59.Rxf8 Kxf8 60.Qf1+ Ke7 61.Bh3 Bd7 62.Be6 Bxe6 63.dxe6 Kxe6 64.Qf8 Ne7 65.Qxd8 Qc8 66.Qxc8+ Nxc8 67.Kc2 Kd7 68.Ne2 Ne7 69.Nc3 Nc6 70.Kb3 Ke6 71.Nd5 Kd7 72.Ka3 Kd8 73.Bb6+ Kd7 74.Nc3 Ke6 75.Ne2 Kd7 76.Ng3 Ne7 77.Kb3 Nc8 78.Nf1 Ne7 79.Ne3 Nc6 80.Kc3 Ne7 81.Kb2 Nc6 82.Kb3 Ke6 83.Kc3 Kd7 84.Nd5 Nb8 85.Kd3 Nc6 86.Nc7 Nb8 87.Kd2 Kc8 88.Kc3 Kd7 89.Kd3 Ke7 90.Nd5+ Kd7 91.Nf6+ Ke6 92.Kc3 Nc6 93.Be3 Ne7 94.Kc2 Nc6 95.Nd5 Kd7 96.Kb3 Kd8 97.Bb6+ Kd7 98.Nc7 Nd4+ 99.Kc3 Nf3 100.Kd3 
  White is clearly winning: +- (4.75)  Depth: 73/95   18:26:24  1448556MN, tb=910620506

On my computer both Leelenstein and Allie play Bf1 and seem to hold it so not sure why it didn't play winning move at tcec. Perhaps due to combination of what was in cache and move history... If anyone else can test this it would be good.

Code: Select all

AllieStein v0.7_dev-net_14.3 - Stockfish 202005232210 1/2-1/2, TCEC Season 18 - Division P 2020
2nbbrk1/3q4/p2p2p1/Pp1Pp1Pp/1P2P2P/4B1N1/4B3/1KR4Q w - - 0 1
Analysis by Brainfish 280520 64 POPCNT:
54.Qh2 Ne7 55.Bf1 Qb7 56.Bh3 Bd7 57.Qg2 Kf7 58.Rf1+ Kg7 59.Rxf8 Kxf8 60.Be6 Bxe6 61.dxe6 Nc6 62.Qf3+ Kg8 63.Bb6 Qe7 64.Qb3 Kh7 65.Ne2 Qe8 66.Nc3 Kg7 67.Kb2 Qe7 68.Ka2 Bxb6 69.axb6 Qf8 70.Nd5 Qf2+ 71.Ka3 Qd2 72.Qc3 Qxc3+ 73.Nxc3 Kf8 74.Nd5 Nb8 75.Kb2 Nc6 76.b7 Nb8 77.Kc3 Ke8 78.Kb3 Kd8 79.Nf6 Nc6 80.Nd7 Ke7 81.b8Q Nxb8 82.Nxb8 Kxe6 83.Nc6 Kd7 84.Na5 Kd8 85.Kc2 Ke8 86.Kc3 Kd7 87.Nb3 Ke6 88.Kd2 Ke7 89.Ke3 Kd7 90.Kd3 Ke7 91.Kd2 Kd7 92.Ke3 Kc6 93.Kd3 Kd7 94.Nd2 Ke7 95.Nf1 Ke6 96.Ne3 Kd7 97.Nd5 Kd8 98.Kc2 Ke8 99.Kd2 Kd7 100.Kc3 Kd8 101.Ne3 Ke7 102.Nf1 Kd7 103.Nd2 Kc6 
  White has a decisive advantage: +- (3.28 ++)  Depth: 73/27   02:32:30  200374MN, tb=206803677

Code: Select all

AllieStein v0.7_dev-net_14.3 - Stockfish 202005232210 1/2-1/2, TCEC Season 18 - Division P 2020
2nbbrk1/8/p2p2p1/Pp1Pp1Pp/1P2P1qP/4B1N1/8/1KR2B1Q w - - 0 1

Analysis by Brainfish 280520 64 POPCNT:

55.Qg2 Ne7 56.Be2 Qd7 57.Qh2 Qb7 58.Qg1 Bc7 59.Bd2 Bd8 60.Kb2 Kg7 61.Qe3 Rf4 62.Bf3 Ng8 63.Bd1 Ne7 64.Qg1 Rf8 65.Rc2 Kg8 66.Be2 Qd7 67.Rc3 Kg7 68.Qd1 Kh7 69.Be3 Kg8 70.Qh1 Bf7 71.Rc2 Be8 72.Rc1 Kg7 73.Qh2 Bc7 74.Bf1 Qd8 75.Qc2 Bb8 76.Bh3 Bd7 77.Qg2 Qe8 78.Ne2 Nc8 79.Qg3 Qf7 80.Bg1 Bxh3 81.Qxh3 Ba7 82.Bxa7 Nxa7 83.Nc3 Qe8 84.Nb1 Rf2+ 85.Ka3 Rf7 86.Qe6 Rf3+ 87.Kb2 Rf2+ 88.Ka1 Qd8 89.Nd2 Rf4 90.Qh3 Rf7 91.Kb2 Rb7 92.Qe6 Qe7 93.Rc3 Qd8 94.Kb3 Qf8 95.Rc1 Qf2 96.Rc2 Qf8 97.Kb2 Qd8 98.Nf1 Qf8 99.Ng3 Rf7 100.Rc1 Qe7 
  White has a decisive advantage: +- (2.66)  Depth: 76/105   17:09:17  1300599MN, tb=1422967112

[jjoshua2]

Code: Select all

AllieStein v0.7_dev-net_14.3 - Komodo 2551.00 Line, TCEC Season 18 - Division P 2020
1r6/4k3/pr1q4/2p1p1p1/2P1Bp1p/RP1PnP1P/Q5P1/2R4K w - - 0 1

Analysis by Stockfish 260520 64 POPCNT:

46.Ra4 Kf6 
  White is clearly better: +- (2.18 ++)  Depth: 66/83   01:45:27  141588MN, tb=79583634
  
46.Ra5 Kf8 47.Rg1 Rc8 48.Ra1 Kg7 49.Rc1 Kf6 50.Qa1 Rcb8 51.b4 Rxb4 52.d4 Qxd4 53.Rxa6+ R4b6 54.Rxb6+ Rxb6 55.Qa8 Qd6 56.Qg8 Ke7 57.Ra1 Ra6 58.Qg7+ Kd8 59.Rg1 Rb6 60.Qxg5+ Qf6 61.Qg8+ Ke7 62.Bd5 Qg6 63.Qc8 Qd6 64.Qh8 Kd7 65.Ra1 Ra6 66.Qh7+ Ke8 67.Qg8+ Kd7 68.Rb1 Rb6 69.Qg7+ Ke8 70.Ra1 Ra6 71.Rg1 Rb6 72.Qf7+ Kd8 73.Qg8+ Kd7 74.Be4 Qf6 75.Ra1 Ra6 76.Qh7+ Ke8 77.Rxa6 Qxa6 78.Bg6+ Kd8 79.Qxh4+ Kd7 80.Qh7+ Kd6 81.Bd3 Qc8 82.h4 Qe6 83.Qh8 Qe7 84.Kh2 Kd7 85.Kg1 Qd8 86.Qh6 Ke7 87.Qg6 Qd4 88.Kh2 Qd6 89.Qxd6+ Kxd6 90.h5 Ke6 91.h6 Kf6 92.g4 fxg3+ 93.Kxg3 Nd1 
  White has a decisive advantage: +- (2.62)  Depth: 67/106   04:31:28  365460MN, tb=259247594

After playing the line with a deep analyses at each point, the deepest analyses here:

Code: Select all

1r3k2/8/pr1q4/R1p1p1p1/2P1Bp1p/1P1PnP1P/Q5P1/6RK b - - 0 1

Analysis by Stockfish 260520 64 POPCNT:

47...Rc8 48.Rc1 Rcb8 49.Kh2 Ke7 50.Rg1 Rc8 51.Re1 Kf6 52.Rc1 Rcb8 53.Rg1 Rc8 54.Kh1 Kg7 55.Rc1 Kf6 56.Qa1 Rcb8 57.b4 Rxb4 58.d4 Qxd4 59.Rxa6+ R4b6 60.Rxb6+ Rxb6 61.Qa8 Qd6 62.Qg8 Ke7 63.Ra1 Ra6 64.Qg7+ Kd8 65.Rg1 Rb6 66.Qxg5+ Qf6 67.Qg8+ Ke7 68.Bd5 Qg6 69.Qc8 Qd6 70.Qh8 Kd7 71.Ra1 Ra6 72.Qg7+ Ke8 73.Rg1 Rb6 74.Be4 Kd8 75.Ra1 Ra6 76.Qg8+ Ke7 77.Qh7+ Ke8 78.Rg1 Rb6 79.Qh5+ Kd7 80.Ra1 Ra6 81.Rc1 Ra2 82.Qh7+ Kc8 83.Bb7+ Kd8 84.Bd5 Rd2 85.Ra1 Rd1+ 86.Rxd1 Nxd1 87.Qxh4+ Qe7 88.Qh6 Nc3 89.Kg1 Nxd5 90.cxd5 c4 91.d6 Qf7 92.Qh8+ Qe8 93.Qh7 Qc6 94.Qe7+ Kc8 95.Kf1 c3 96.d7+ Qxd7 
  White has a decisive advantage: +- (2.92)  Depth: 76/121   09:00:01  1047370MN, tb=1140509785

Analysed a lot of similar positions deeply and they all are about the same with close to +3 and a lot of shuffling.

Allie vs Ethereal and even Allie vs Stoofvlees were analyzed by Occy and also have missed wins. So that's up to 5. There was even a possible win in the other SF game that has not been deeply analyzed.

[Dann Corbit]

It is not solely a NN problem to miss wins. I have seen AB engines up a full piece and then not win the game. In fact, I guess that most losses or draws, somehow or somewhere, contain a missed win. If we could see out 1000 plies tactically, we could perhaps exploit an absurd number of missed win possibilities. Similarly, if we could formulate a strategic plan with a structural devastation that gave us enormous space, development and mobility advantages, we might win many games that were otherwise lost or drawn.

That is on the one hand. On the other hand, I think that you have seen what many have been talking about recently. The much more shallow searches become a bigger and bigger disadvantage when the board gets sparse and so the NN engines really do have more trouble in the endgame.

[jjoshua2]

Gaining depth extremely slowly, because of fail lows probably, so going to stop here without resolving to mate scores.

Code: Select all

Fire 8_beta - AllieStein v0.7_dev-net_14.3 Line, TCEC Season 18 - Division P 2020
1b6/1P4p1/1nP2k2/1P3p2/2B1pP2/6P1/6K1/8 w - - 0 1

Analysis by Brainfish 280520 64 POPCNT:

48.Bg8 g5 
  Black is clearly winning: -+ (-22.89 --)  Depth: 58/62   29:16:18  1723536MN, tb=2147483647