AB and NN engines

[Fritz 0]

It seems that there is a generally accepted opinion that NN engines play in a more "human-like" style than traditional AB engines. But it doesn't really make sense to me. While the evaluation of NN engines is based on the games that an engine plays against itself, the evaluation of AB engines is based on parameters set by a human (in the case of Komodo, a Grandmaster). Therefore, how the former can be more "human-like" than the latter?

[Sopel]

NN engines are not more "human-like". MCTS engine could be maybe considered so but there is no evidence that people can distinguish play of an MCTS engine and AB engine, let alone find it "human-like".

[Uri Blass]

I totally agree that NN engines are not more human like.
NN engines have better evaluations but not because their evaluation is more similiar to humans.

NN engines are designed to play better chess and not to emulate humans.
They are designed to have knowledge in the evaluation that humans do not have so humans often see scores that they do not understand and not because of search(there are many positions when stockfish suggest that one side is winning or that it is a draw when non NN engines and humans usually do not see it because of inferior evaluation).

Human often also have a wrong knowledge based on what they learn(for example that they have an advantage because of pair of bishops even in position that it is not relevant) and here is one example

This is from a game from 1905 when black was one of the best players at that time
Aron Nimzowitsch

Black refused to a draw offer when he claimed that the fact that he has a pair of bishops gives him some advantage even if the advantage is not enough to win the game.

[fen]3b4/3b1k2/1p1p4/p1pNp1p1/P1P1P1P1/1PB2PK1/8/8 w - - 5 1 [/fen]

Stockfish correctly as white does not think that black is better.

The next moves of the game were Bxe5 dxe5 Nxb6 Bxb6 and black has to agree to a draw because he understood the bishops are useless.

Unfortunately stockfish15 show some advantage for black after these moves but it is not a decisive advantage and stockfish has a big fail high for white after these moves(do not know based on what exactly because it cannot see 50 plies forward but the score is only 0.74 for black after some minutes of search that does not suggest a winning evaluation).

[fen]8/3b1k2/1b6/p1p1p1p1/P1P1P1P1/1P3PK1/8/8 w - - 0 3[/fen]

Stockfish_15_x64_avx2:
NNUE evaluation using nn-6877cd24400e.nnue enabled

35/41 00:34 505,528k 14,630k -1.79 Kg3-f2 Kf7-e7 Kf2-e1 Bd7-c8 Ke1-e2 Bb6-d8 Ke2-d2 Ke7-f6 Kd2-e1 Kf6-f7 Ke1-d2 Bd8-e7 Kd2-e3 Bc8-b7 Ke3-e2 Kf7-e6 Ke2-d2 Ke6-d7 Kd2-e1 Bb7-d5 Ke1-d1 Be7-f6 Kd1-e2 Kd7-e8 Ke2-e3 Bf6-e7 e4xd5 Ke8-f7 Ke3-e4 Kf7-f6 d5-d6 Be7xd6 Ke4-d5 Kf6-e7 Kd5-c6 Ke7-e6 Kc6-b5 Ke6-d7 Kb5-b6 Bd6-e7 Kb6xa5 Kd7-c6 Ka5-a6
36/35+ 00:35 511,806k 14,622k -1.74 Kg3-f2
36/37+ 00:44 654,259k 14,814k -1.62 Kg3-f2
36/41+ 01:55 1,698,397k 14,669k -0.74 Kg3-f2
36/41 02:08 1,914,280k 14,848k -0.56 Kg3-f2 Bb6-c7 Kf2-e2 Bd7-e6 Ke2-e1 Bc7-d8 Ke1-d2 Kf7-e8 Kd2-e1 Ke8-e7 Ke1-d1 Ke7-d6 Kd1-e2 Be6-f7 Ke2-d1 Bd8-f6 Kd1-c2 Bf6-g7 Kc2-d2 Bf7-g8 Kd2-e1 Bg7-f6 Ke1-d1 Kd6-e7 Kd1-d2 Bg8-e6 Kd2-e2 Be6-d7 Ke2-d1 Bd7-c6 Kd1-c1 Ke7-d7 Kc1-b1 Bc6xe4+ f3xe4
37/46- 02:11 1,948,205k 14,829k -0.75 Kg3-f2 Bb6-c7
37/46 02:11 1,955,188k 14,831k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Ke6-d6 Kd1-e1 Bc8-e6 Ke1-d2 Be6-g8 Kd2-d1 Bd8-f6 Kd1-d2 Bf6-g7 Kd2-c2 Kd6-e7 Kc2-b1 Ke7-e8 Kb1-c2 Ke8-d8 Kc2-d1 Kd8-d7 Kd1-c1 Bg7-f8 Kc1-d2 Kd7-e7 Kd2-c2 Bg8-h7 Kc2-d1 Bh7-g6 Kd1-d2 Ke7-f6 Kd2-d1 Kf6-f7 Kd1-c1 Kf7-g8
38/39 02:12 1,962,568k 14,830k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Ke6-d6 Kd1-e1 Bc8-e6 Ke1-d2 Be6-f7 Kd2-c1 Bf7-g6 Kc1-d2 Bg6-h7 Kd2-d1 Bd8-f6 Kd1-e2 Bh7-g8 Ke2-d1 Kd6-e7 Kd1-d2 Bg8-h7 Kd2-c1 Bf6-g7 Kc1-d2 Ke7-f7 Kd2-d1 Bh7-g8 Kd1-e2 Kf7-e7 Ke2-d2 Bg7-f6 Kd2-c1 Ke7-d7 Kc1-d1
39/41 02:13 1,977,133k 14,822k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Ke6-d6 Kd1-e1 Bc8-e6 Ke1-d2 Be6-f7 Kd2-c1 Bf7-g6 Kc1-c2 Kd6-e6 Kc2-d2 Ke6-f7 Kd2-d1 Bd8-b6 Kd1-e1 Kf7-g8 Ke1-d2 Bb6-d8 Kd2-d1 Kg8-h7 Kd1-c2 Bd8-e7 Kc2-d1 Bg6-e8 Kd1-c2 Kh7-g8 Kc2-d1 Be8-d7 Kd1-c1 Be7-d8 Kc1-b1 Bd8-f6
40/40 02:13 1,979,102k 14,821k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-d2 Bf6-g7 Kd2-e2 Kd8-e7 Ke2-d1 Bg7-h8 Kd1-c1 Ke7-f7 Kc1-b2 Bc8-e6 Kb2-b1 Bh8-g7 Kb1-c2 Be6-c8 Kc2-d2 Bc8-b7 Kd2-e2 Bg7-f8 Ke2-d2 Kf7-g8 Kd2-e1 Bf8-d6 Ke1-e2 Bb7-c8 Ke2-e1 Kg8-f7
41/43 02:14 1,989,914k 14,813k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-d2 Bf6-g7 Kd2-c1 Bc8-b7 Kc1-b2 Bb7-c6 Kb2-c2 Kd8-e7 Kc2-d1 Bc6-d7 Kd1-c2 Bd7-e6 Kc2-d1 Ke7-f6 Kd1-e2 Bg7-f8 Ke2-d1 Be6-g8 Kd1-c2 Bg8-h7 Kc2-b1 Kf6-e7 Kb1-a1 Bf8-g7 Ka1-a2 Bh7-g6 Ka2-b2 Bg6-e8
42/45 02:14 1,997,342k 14,807k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-d2 Bf6-g7 Kd2-c1 Bc8-b7 Kc1-b2 Bb7-c6 Kb2-c1 Kd8-c7 Kc1-b2 Bc6-d7 Kb2-a2 Kc7-d8 Ka2-a1 Bd7-c8 Ka1-a2 Bg7-f8 Ka2-b2 Kd8-e7 Kb2-c2 Bf8-g7 Kc2-d2 Bc8-e6 Kd2-e2 Be6-d7 Ke2-e1 Bg7-f6 Ke1-d1 Bd7-c6 Kd1-e2 Ke7-f7
43/46 02:16 2,018,824k 14,807k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-d2 Bf6-g7 Kd2-c1 Bc8-b7 Kc1-b2 Bb7-c6 Kb2-c1 Kd8-c7 Kc1-d1 Bg7-f6 Kd1-c1 Bf6-e7 Kc1-d1 Kc7-d8 Kd1-e1 Bc6-d7 Ke1-e2 Be7-f6 Ke2-e1 Bd7-e6 Ke1-e2 Kd8-e7 Ke2-e1 Be6-g8 Ke1-f2 Bg8-h7 Kf2-e2 Bh7-f5 Ke2-e1
44/43 02:18 2,047,731k 14,785k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-e2 Bc8-e6 Ke2-e1 Bf6-e7 Ke1-d1 Kd8-e8 Kd1-c2 Ke8-f7 Kc2-c1 Be7-f8 Kc1-d1 Kf7-g8 Kd1-c2 Kg8-h7 Kc2-d2 Bf8-g7 Kd2-c1 Kh7-g8 Kc1-d2 Kg8-f8 Kd2-e2 Kf8-e7 Ke2-e3 Bg7-f8 Ke3-d3 Ke7-d7 Kd3-c2 Be6-f5
45/44 02:23 2,125,997k 14,807k -0.75 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-e2 Bc8-e6 Ke2-e1 Bf6-e7 Ke1-d1 Be6-g8 Kd1-e2 Be7-f8 Ke2-d1 Bg8-f7 Kd1-c1 Kd8-c8 Kc1-b1 Kc8-d7 Kb1-a1 Bf7-g6 Ka1-b2 Kd7-c6 Kb2-c1 Bg6-f5 Kc1-c2 Bf5-e6 Kc2-c1 Kc6-d7 Kc1-d1 Be6-f5 Kd1-d2 Kd7-d6 Kd2-e1
46/49 02:44 2,552,936k 15,495k -0.74 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d8 Ke2-e1 Kf7-e6 Ke1-d1 Bd8-f6 Kd1-d2 Ke6-d7 Kd2-e1 Kd7-d8 Ke1-d1 Bf6-g7 Kd1-c2 Kd8-e7 Kc2-d1 Ke7-f6 Kd1-c2 Bc8-e6 Kc2-b1 Kf6-f7 Kb1-a2 Kf7-g8 Ka2-b2 Be6-c8 Kb2-c2 Kg8-h8 Kc2-d2 Bc8-d7 Kd2-c1 Bg7-f6 Kc1-d2 Bf6-e7 Kd2-c2 Kh8-h7 Kc2-d2 Kh7-g8 Kd2-c2 Be7-f8 Kc2-b2 Kg8-h8 Kb2-c1 Bd7-e6
47/49 02:52 2,742,587k 15,878k -0.74 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d6 Ke2-d2 Kf7-e7 Kd2-c2 Ke7-d8 Kc2-b2 Bd6-f8 Kb2-a2 Bc8-e6 Ka2-b2 Be6-d7 Kb2-a2 Bd7-e8 Ka2-b1 Kd8-d7 Kb1-c1 Bf8-e7 Kc1-b1 Be8-f7 Kb1-c2 Be7-f6 Kc2-b1 Kd7-d8 Kb1-c1 Bf7-g6 Kc1-d1 Bg6-h7 Kd1-c2 Kd8-e7 Kc2-b2 Ke7-f8 Kb2-c1 Kf8-g8 Kc1-d1 Bf6-d8 Kd1-c2 Bh7-g6 Kc2-d3 Bd8-c7 Kd3-d2 Bc7-d6 Kd2-d1
48/48 02:56 2,834,493k 16,043k -0.74 Kg3-f2 Bb6-c7 Kf2-f1 Bd7-c8 Kf1-e2 Bc7-d6 Ke2-d2 Bc8-b7 Kd2-c1 Bb7-c6 Kc1-d1 Kf7-f6 Kd1-d2 Bc6-e8 Kd2-d1 Bd6-f8 Kd1-e1 Kf6-g6 Ke1-e2 Kg6-h7 Ke2-d1 Kh7-h8 Kd1-c2 Bf8-g7 Kc2-d1 Kh8-g8 Kd1-e1 Be8-c6 Ke1-e2 Kg8-h7 Ke2-d2 Bc6-b7 Kd2-c2 Bb7-c8 Kc2-d1 Kh7-g8 Kd1-e2 Bg7-f6 Ke2-d2 Kg8-h7 Kd2-e2 Bc8-d7 Ke2-d2 Kh7-h8 Kd2-c1 Bf6-e7 Kc1-b2 Bd7-e8

[Fritz 0]

Incredible position! Black is two bishops up, but none of these bishops is worth more than a pawn.

[hgm]

It should not be so difficult to recognize such positions in a conventional evaluation. One could tabulate pawn structures that cause impenetrable barriers to various types of pieces in a hash table, and consult that table when evaluating a pawn chain.

Of course no one will ever do that, because there is no Elo in it.

[Uri Blass]

Incredible position! Black is two bishops up, but none of these bishops is worth more than a pawn.

I composed another position when black is even 2 bishops and 2 pawns up but does not win.
[fen]8/3b1kp1/pb4p1/p1p1p1p1/P1P1PpP1/1P3PK1/8/8 w - - 0 3[/fen]

[Ajedrecista]

Hello Branislav:

Incredible position! Black is two bishops up, but none of these bishops is worth more than a pawn.

Indeed! I was unaware of this position. It looks like the full game record is unavailable according to what I found online. Thanks to Uri's hints of the year and the player with black pieces, I found that the game was Neumann-Nimzowitsch, Vienna (1905), round 17, 6th March. Moves were 51st and 52nd:

https://nimzowitsch.net/wien-1905.html
https://nimzowitsch.net/wien-1905-parti ... d-265.html

Few additional info can be found. The game started with Spanish Opening according to a newspaper of the following day:

Neue Freie Presse, 7. März 1905
Neue Freie Presse, 7. März 1905, Seite 8

The mention to the Spanish Opening (Spanische Eröffnung) is at the middle of the first column of page 8 of the issue, you only have to search a few lines before 51. Lc3Xe5 d6Xe5.

Other source to look into is the page 25 of Ludwig Bachmann's Schachjahrbuch für 1905, I. Teil.

Further info is welcome.

Regards from Spain.

Ajedrecista.

[Raphexon]

It seems that there is a generally accepted opinion that NN engines play in a more "human-like" style than traditional AB engines. But it doesn't really make sense to me. While the evaluation of NN engines is based on the games that an engine plays against itself, the evaluation of AB engines is based on parameters set by a human (in the case of Komodo, a Grandmaster). Therefore, how the former can be more "human-like" than the latter?

AB is a search method.
NNs are used for eval. (but also move ordering in Leela's case)

They aren't exclusionary...

[AndrewGrant]

It seems that there is a generally accepted opinion that NN engines play in a more "human-like" style than traditional AB engines.

Maybe. I think the claim is laughable. The number of players capable of understanding the choices of engines is very few. And of those, ever fewer care to make commentary on computer chess.

People see what they want to see. A universal truth almost. Far beyond the game of chess.

[Fritz 0]

I have such impression too, but I am a fairly weak player, slightly above 1900 chess.com, so maybe I don't see what stronger players do. I've played much more AB than NN engines, but I've only noticed that Dragon levels are perhaps a skill level stronger than corresponding Komodo levels. In playing style, I haven't noticed any difference.