Why SF missed 16.g4 in Game 12 of S21 TCEC Superfinal

186 views
Skip to first unread message

Bryan

unread,
Jul 22, 2021, 3:50:27 PMJul 22
to FishCooking

In Game 12 of S21 of TCEC Superfinal, GM Sadler noticed that SF as white had missed 16.g2-g4, and Aloril confirmed (to 100+Gn) that it was over +3. Using the same SFdev playing in the superfinal I decided to try to understand why SF missed 16.g2-g4. So I did a search on one thread at move 16 black to move after 16.g2-g4.

Move 16: black to move after 16.g4: hash cleared: multipv=1: threads=1
FEN: r3r2k/p1N2p1P/2p1bnp1/2P2p2/5Np1/8/PPP5/2KR3R w - - 0 26
Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
[...]
 35/56    01:04     59,779k    924k    -0.59    h5xg4 h4-h5 e6xf5 h5-h6 g7-g6 Nc3xd5 Nf6-e4 h6-h7+ Kg8-h8 Qd2-e3 Qd8-g5 Bb5xc6 b7xc6 Nd5-c7 Bc8-e6 Qe3-d4+ Qg5-f6 Qd4xf6+ Ne4xf6 Rd1-d6 g6-g5 Nf4xe6 f7xe6 Nc7xa8 Re8xa8 Rd6xe6 Ra8-f8 Kc1-d2 Nf6-e4+ Kd2-e1 g4-g3 b2-b4 f5-f4 Re6xe4 f4-f3 Rh1-g1 f3-f2+ Ke1-f1 f2xg1Q+ Kf1xg1 Rf8-f2 Re4-e6 Rf2xc2 Re6xc6 Kh8xh7 Rc6-a6 Rc2-c4 Ra6xa7+ Kh7-g6 b4-b5 Rc4xc5
[...]
 36/64    01:27     82,487k    939k    -0.41    h5xg4 h4-h5 e6xf5 h5-h6 g7-g6 Nc3xd5 Nf6-e4 h6-h7+ Kg8-h8 Qd2-e3 Qd8-g5 Bb5xc6 b7xc6 Nd5-c7 Bc8-e6 Qe3-d4+ Qg5-f6 Qd4xf6+ Ne4xf6 Nf4xe6 f7xe6 Rh1-h6 Ra8-d8 Nc7xe8 Rd8xe8 Kc1-d2 g4-g3 Rd1-h1 Nf6-e4+ Kd2-e2 g6-g5 Ke2-f3 e6-e5 Rh1-e1 g5-g4+ Kf3-g2 Re8-d8 Re1xe4 f5xe4 Rh6-d6 Rd8-e8 Kg2xg3 e4-e3 Rd6-d3 Kh8xh7 Rd3xe3 Kh7-g6 Kg3xg4 Re8-b8 b2-b3 Kg6-f6

At at d=35 the mistake is that SF thinks that at move 26, white's best move is Rd1-d6. However, here white has Nf4xe6. At d=36 the mistake is that at move 29, white to move, SF thinks white's best move is Kc1-d2 (which is a draw), when white's winning move here is Rd1-h1, which restricts black's king to h8 and threatens the indefensible g6 pawn and puts pressure on the black knight on f6.

To find out why SF has difficulty finding 26.Nf4xe6 in that line I did a search.

Move 26: black to move after 16.g4 h5xg4 h4-h5 e6xf5 h5-h6 g7-g6 Nc3xd5 Nf6-e4 h6-h7+ Kg8-h8 Qd2-e3 Qd8-g5 Bb5xc6 b7xc6 Nd5-c7 Bc8-e6 Qe3-d4+ Qg5-f6 Qd4xf6+ Ne4xf6 Nf4xe6: hash cleared: multipv=1: threads=1
FEN: r3r2k/p1N2p1P/2p1Nnp1/2P2p2/6p1/8/PPP5/2KR3R b - - 0 26
Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
[...]
 30/46    00:20     20,366k    1,011k    -0.39    f7xe6 Rh1-h6 Ra8-b8 Nc7xe8 Rb8xe8 b2-b4 Nf6-d5 c2-c4 Nd5-e3 Rd1-h1 Ne3xc4 Rh6xg6 Nc4-e5 Rg6-g5 Ne5-d3+ Kc1-d2 Nd3xb4 Rh1-h6 Nb4xa2 Rg5-g8+ Re8xg8 h7xg8Q+ Kh8xg8 Rh6xe6 Na2-b4 Re6-f6 Kg8-g7 Rf6xf5 Kg7-g6 Rf5-f8 Kg6-g7 Rf8-f4 Nb4-d5 Rf4xg4+ Kg7-f7 Rg4-g3
 
Here SF thinks that at move 29 white's best move is b2-b4. But b2-b4 is drawish. White's winning move here is, again, Rd1-h1. The move Rd1-h1 is in the pv here, but two moves later. And recall that at the d=36 iteration where SF thought 29.Kc1-d2 was white's best move, SF was also failing to see 29.Rd1-h1.

So the mistakes at d=35 and at d=36 are both on account of SF failing to see 29.Rd1-h1. To find out why SF has difficulty seeing this 29.Rd1-h1, I did a search at Move 29, black to move, after playing 29.Rd1-h1.

Move 29: black to move after [see previous]: hash cleared: multipv=1: threads=1
FEN: 4r2k/p6P/2p1pnpR/2P2p2/6p1/8/PPP5/2K4R b - - 1 29
Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
[...]
 18/28    00:00     438k    997k    -0.64    Nf6-h5 Rh6xg6 Nh5-f4 Rg6-g5 Nf4-h3 Rg5-g8+ Re8xg8 h7xg8N Kh8xg8 b2-b4 a7-a6 a2-a4 Nh3-f2 b4-b5 a6xb5 Rh1-f1 b5xa4 Rf1xf2 Kg8-f7 Kc1-d1 g4-g3 Rf2-d2 a4-a3 c2-c4 f5-f4 Kd1-e2
 
Obviously this search is very shallow, but it allows us to find the root of the initial blindspot. The mistake here is that at move 37, SF thinks white's best move is Rh1-f1, but white has a tablebase win with 37.a4xb5. Even 37.Rh1-f1 b5xa4 Rf1xf2 is (likely) a tablebase win. In this line at move 41 SF thinks white's best move is c2-c4, but that appears to be a draw. There white has two moves that appear to be tablebase wins: Kd1-e1 and Kd1-e2.

Now I went back to move 16 and did the search again, this time with threads=15.

Move 16: black to move after 16.g4: hash cleared: multipv=1: threads=1
FEN: r1bqr1k1/pp3pp1/2n1pn2/1BPp1P1p/5NPP/2N5/PPPQ4/2KR3R b - g3 0 16
Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
[...]
 33/63    00:19     189,548k    9,761k    -0.37    h5xg4 h4-h5 e6xf5 h5-h6 g7-g6 Nc3xd5 Nf6-e4 h6-h7+ Kg8-h8 Qd2-e3 Qd8-g5 Nd5-c7 Bc8-e6 Bb5xc6 b7xc6 Qe3-d4+ Qg5-f6 Qd4xf6+ Ne4xf6 Nf4xe6 f7xe6 Rh1-h6 Ra8-c8 Nc7xe8 Rc8xe8 b2-b4 Nf6-d5 c2-c4 Nd5-e3 Rd1-h1 Ne3xc4 Rh6xg6 Nc4-e5 Rg6-g5 Ne5-d3+ Kc1-d2 Nd3-f2 Rg5-g8+ Re8xg8 h7xg8Q+ Kh8xg8 Rh1-c1 a7-a6 a2-a4 Kg8-f7 b4-b5 a6xb5 Kd2-e3 Nf2-h3 a4xb5 f5-f4+ Ke3-e2 c6xb5

Again, the mistake is at move 29, where SF thinks 29.b2-b4 is white's best move, failing to see 29.Rd1-h1.
Reply all
Reply to author
Forward
0 new messages