Re: LC0's Superior Performance in Unconventional Chess Positions

[urib...@gmail.com]

one possible reason may be that stockfish's evaluation is weak in these positions
I do not know what are the positions that you tried but
I found stockfish does not know the simple principle of trading material when it has a very big advantage

some analysis with stockfish
The evaluation does not go up from depth 1

The reason is obvious stupid stockfish consider trading material as bad for white so it does not extend these lines

Of course it may win in a practical game but only after many moves and normal chess engines are going to win faster with white against it


FEN: 1nb1kbn1/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQ - 0 1

Stockfish_23070621_x64_avx2:
NNUE evaluation using nn-c38c3d8d3920.nnue enabled
 1/1 00:00  104 52k +14.08 e2-e3
 2/2 00:00  1k 635k +14.01 Ng1-f3 e7-e6
 3/3 00:00  2k 824k +14.08 e2-e3
 4/3 00:00  2k 934k +14.03 Ng1-f3 e7-e6
 5/5 00:00  3k 1,695k +13.71 Ng1-f3 e7-e6 Nb1-c3 Nb8-c6
 6/6 00:00  5k 2,292k +13.81 Nb1-c3 Nb8-c6 Ng1-f3
 7/6 00:00  7k 3,749k +13.71 Ng1-f3 Ng8-f6 e2-e3 Nb8-c6 Nb1-c3
 8/8 00:00  13k 4,290k +13.71 Nb1-c3 Ng8-f6 Ng1-f3 Nb8-c6 d2-d3 e7-e6
 9/7 00:00  48k 8,017k +13.58 Ng1-f3 Nb8-c6 b2-b3 Ng8-f6 Nb1-c3 e7-e6 e2-e4
 10/10 00:00  66k 9,433k +13.70 Nb1-c3 d7-d6 Nc3-b5 Ke8-d8 Ng1-f3 Ng8-f6
 11/11 00:00  117k 9,710k +13.63 Nb1-a3 Ng8-f6 Na3-b5 Ke8-d8 f2-f4 Nb8-c6 Ng1-f3
 12/12 00:00  177k 9,296k +13.65 Nb1-a3 a7-a6 e2-e3 e7-e6 Na3-c4 Nb8-c6 Ng1-e2 Bf8-d6
 13/13 00:00  330k 9,432k +13.64 Nb1-a3 Nb8-c6 Na3-b5 Ke8-d8 Ng1-f3 a7-a6 Nb5-a3 e7-e6 Na3-c4 Bf8-b4
 14/16 00:00  520k 9,449k +13.53 Nb1-a3 a7-a6 d2-d3 e7-e6 Na3-c4 d7-d5
 15/17 00:00  1,560k 11,643k +13.46 Nb1-c3 Nb8-c6 d2-d4 e7-e6 a2-a3 b7-b6 Nc3-b5 Bf8-d6 d4-d5 e6xd5 Qd1xd5 Ke8-f8
 16/20 00:00  2,256k 12,062k +13.49 Nb1-c3 Nb8-c6 d2-d4 e7-e6 a2-a3 Ng8-f6 e2-e4 b7-b6 f2-f3 Bc8-b7 Ke1-f2 Bf8-e7 Bc1-f4 Be7-d6 e4-e5
 17/25 00:00  2,988k 11,905k +13.49 Nb1-c3 Nb8-c6 d2-d4 e7-e6 a2-a3 Ng8-f6 e2-e3 Nf6-d5 Nc3-e2 Nd5-f6 Ng1-f3 Bf8-d6 c2-c4 Ke8-f8 b2-b4 b7-b6 c4-c5 b6xc5
 18/21 00:00  3,691k 11,947k +13.51 Nb1-c3 Ng8-f6 e2-e3 Nb8-c6 d2-d4 e7-e6 a2-a3 Nf6-d5 Nc3-e4 Nd5-f6 Ne4-d2 e6-e5 c2-c3 e5xd4 c3xd4 Bf8-d6
 19/25 00:00  7,322k 11,984k +13.60 g2-g3 Ng8-f6 e2-e3 Nb8-c6 Bf1-g2 e7-e6 a2-a3 Bf8-d6 Ng1-e2 b7-b6 d2-d3 Ke8-f8 O-O Kf8-g8 c2-c4 Nf6-g4 f2-f4 g7-g6 Nb1-c3
 20/27 00:00  8,048k 11,976k +13.63 g2-g3 Ng8-f6 Bf1-g2 Nb8-c6 e2-e3 e7-e6 a2-a3 Bf8-d6 Ng1-e2 b7-b6 d2-d3 g7-g6 Nb1-d2 Bc8-a6 c2-c4
 21/19 00:00  8,122k 11,979k +13.63 g2-g3 Ng8-f6 Bf1-g2 Nb8-c6 e2-e3 e7-e6 Ng1-e2 Bf8-d6 a2-a3 Ke8-f8 d2-d3 g7-g6 Nb1-c3 Nc6-a5 b2-b4
 22/23 00:00  8,888k 11,882k +13.56 g2-g3 Ng8-f6 Bf1-g2 b7-b6 e2-e3 e7-e6 Ng1-e2 Nb8-c6 a2-a3 Bf8-d6 O-O a7-a5 d2-d3 Bc8-a6 Nb1-d2 g7-g6 c2-c4 Ke8-f8 Ra1-a2 Kf8-g8
 23/25 00:00  9,538k 11,819k +13.60 g2-g3 Ng8-f6 Bf1-g2 Nb8-c6 e2-e3 e7-e6 d2-d3 Bf8-b4+ c2-c3 Bb4-d6 Ng1-e2 a7-a5 O-O b7-b6 c3-c4 Ke8-f8 Nb1-c3
 24/22 00:00  9,931k 11,795k +13.60 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 Ng8-f6 a2-a3 Bf8-d6 Ng1-e2 Nb8-c6 O-O Ke8-f8 d2-d3 a7-a5 Nb1-d2 Kf8-g8 c2-c4 g7-g6 d3-d4 Bc8-a6
 25/28 00:00  11,640k 11,722k +13.63 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 Ng8-f6 a2-a3 Bf8-d6 Ng1-e2 Nb8-c6 O-O Ke8-f8 d2-d3 Bc8-a6 f2-f4 Kf8-g8 Nb1-c3 Nc6-e7 e3-e4 g7-g6
 26/23 00:01  13,384k 11,679k +13.60 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 Ng8-f6 a2-a3 Bf8-d6 Ng1-e2 Ke8-f8 d2-d3 Nb8-c6 O-O a7-a5 c2-c4 Bc8-a6 Nb1-c3 g7-g6 f2-f4 Kf8-g8 Rf1-e1 Bd6-f8 e3-e4
 27/30 00:01  13,736k 11,660k +13.63 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 Ng8-f6 Ng1-e2 Nb8-c6 a2-a3 Bf8-d6 d2-d3 Ke8-f8 O-O a7-a5 c2-c4 g7-g6 Nb1-d2 Kf8-g8 d3-d4 Bc8-b7 Rf1-e1 Bd6-e7 e3-e4 e6-e5
 28/35 00:01  22,998k 11,674k +13.58 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 a7-a5 Ng1-e2 Bc8-a6 d2-d3 Nb8-c6 a2-a3 g7-g6 Nb1-d2 Ng8-e7 O-O Ne7-f5 c2-c4 Nf5-d6 Qd1-b3 Bf8-g7 f2-f4 Ke8-f8
 29/33 00:02  32,393k 11,540k +13.58 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 a7-a5 Ng1-e2 Bc8-a6 d2-d3 Nb8-c6 a2-a3 Bf8-d6 O-O Ng8-f6 Nb1-d2 Ke8-f8 c2-c4 b6-b5 b2-b3 b5xc4 b3xc4 Bd6-c5 Qd1-a4 Bc5-a7 Bc1-b2 Ba7-b6 e3-e4 Kf8-g8
 30/35 00:03  38,418k 11,407k +13.60 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 a7-a5 Ng1-e2 Bc8-a6 d2-d3 Nb8-c6 O-O Ng8-f6 a2-a3 Bf8-d6 Rf1-e1 Ke8-f8 h2-h4 g7-g6 Nb1-d2 Kf8-g8 c2-c4 b6-b5 b2-b3 b5xc4 b3xc4 Kg8-g7 d3-d4 Ba6-b7
 31/35 00:04  51,330k 11,465k +13.56 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 Nb8-c6 Ng1-e2 Ng8-e7 O-O Ne7-f5 d2-d3 Bc8-a6 Nb1-d2 g7-g6 a2-a3 Bf8-g7 g3-g4 Nf5-d6 f2-f4 Ke8-f8 Nd2-f3 Kf8-g8 Ra1-b1 f7-f5 h2-h3 Nc6-e7 g4xf5 g6xf5
 32/35 00:05  57,840k 11,476k +13.58 g2-g3 e7-e6 e2-e3 b7-b6 Bf1-g2 Nb8-c6 d2-d3 g7-g6 Ng1-e2 Ng8-e7 O-O Ne7-f5 f2-f4 Bc8-a6 g3-g4 Nf5-d6 Rf1-f3 Bf8-g7 Rf3-h3 h7-h6 Nb1-d2 Ke8-f8 Nd2-f3 Nd6-b7 Rh3-h4 Kf8-g8 a2-a3 Nb7-c5
 33/39 00:14  160,126k 10,929k +13.53 g2-g3 e7-e6 Bf1-g2 b7-b6 Nb1-c3 Nb8-c6 Ng1-h3 Bc8-a6 O-O Ng8-f6 d2-d3 Bf8-e7 e2-e3 Ke8-f8 a2-a3 Kf8-g8 Nc3-e2 g7-g6 b2-b4 Ba6-c8 c2-c4 h7-h5 Ne2-c3 a7-a6 d3-d4 Bc8-b7 f2-f3
 34/41 00:21  237,661k 10,959k +13.53 g2-g3 e7-e6 Bf1-g2 b7-b6 Ng1-h3 Nb8-c6 f2-f3 Ng8-f6 O-O d7-d5 e2-e3 Bf8-d6 Nh3-f2 Ke8-f8 d2-d3 Kf8-g8 a2-a3 Bc8-d7 Nb1-d2 h7-h5 Rf1-e1 g7-g6 d3-d4 Nc6-a5 b2-b3 c7-c5 Bc1-b2 c5-c4 e3-e4 c4xb3
 35/40 00:26  286,623k 11,011k +13.53 g2-g3 Nb8-c6 e2-e3 e7-e6 d2-d3 b7-b6 Ng1-e2 g7-g6 Bf1-h3 Ng8-e7 Nb1-d2 Bc8-b7 O-O Bf8-g7 Rf1-e1 Ne7-f5 g3-g4 Nc6-e5 g4xf5 Ne5-f3+ Nd2xf3 Bb7xf3 f5xg6 f7xg6 a2-a3 Bf3xe2 Qd1xe2
 36/42 00:42  469,569k 11,018k +13.53 g2-g3 Nb8-c6 e2-e3 e7-e6 d2-d3 g7-g6 Ng1-e2 Ng8-e7 Bf1-h3 Ne7-f5 O-O b7-b6 Nb1-d2 Bf8-h6 Nd2-f3 Bc8-b7 c2-c3 a7-a5 Nf3-e1 Ke8-f8 f2-f4 Nf5-d6 e3-e4 Kf8-g8 Ne1-f3 Bh6-g7 a2-a3 Bb7-a6 e4-e5 Nd6-f5 g3-g4 Nf5-e7 Bh3-g2 Ba6-b7 c3-c4
 37/43 00:47  516,563k 10,984k +13.53 g2-g3 Nb8-c6 e2-e3 b7-b6 Ng1-e2 e7-e6 d2-d3 g7-g6 Bf1-h3 Ng8-e7 O-O Ne7-f5 Nb1-d2 Bf8-h6 Nd2-f3 Bc8-b7 c2-c3 a7-a5 Nf3-e1 Ke8-f8 Ra1-b1 Kf8-g8 f2-f4 Nf5-d6 Ne1-c2 f7-f5 a2-a3 Bb7-a6 c3-c4 Nc6-e7 Rf1-f2 Ba6-b7 b2-b3 Bh6-g7 b3-b4 Kg8-f8

ב-יום שישי, 7 ביולי 2023 בשעה 18:49:35 UTC+3, Donald Lin כתב/ה:

I've experimented with challenging lc0 with various unconventional games positions. These include situations where multiple pawn promotions are possible for both sides, entirely arbitrary positions (think about dropping pieces randomly on a board), or even guaranteed unreachable positions where I simply remove an arbitrary number of pieces from the initial chess position and/or some chess variant such as chess960 (fischer). I tried to let lc0 play against Stockfish and expect stockfish to fare better in these unusual scenarios because of the nature of its search, yet LC0 seemed to significantly outperform Stockfish.

I understand that lc0 is trained through a self-play manner, and relies on neural net heavily when making decisions. In terms of these bizarre positions that I input, I believe it's fairly unlikely that they are seen during lc0's training. Why does LC0 still excel under these peculiar circumstances? Does anyone have any thoughts on this? Any insight or suggestions would be greatly appreciated!