AB and NN engines

[Ovyron]

All this talk is subjective, what is human-like to a person can look machine-like to another. In fact, when I lose against +2400 human players they beat me with engine-like moves, that feel like that on the live game, but turned out to not be the very best, anyway...

ABs are more tactical than NNs, so NNs will play moves that are more shallow and are stronger based on evaluation. They feel more human-like because humans can understand the concept more easily, why they're best. While AB's moves can be indecipherable, after the dust is settled you may be the exchange down while the engine says you're up and you can't understand why, or you're in a position of the sharpest nature where a single inaccuracy is fatal, but AB shows high eval because for an engine it's easy to avoid those inaccuracies.

This has been the flaw of engines with the highest rating when used by humans to analyze their games, they suggest moves that lead to positions the humans won't understand and misplay, hurting them instead of helping, specially in the NNUE times, where the winning positions can the most difficult to play even for classical eval!

NN engines (only Leela and its derivatives) don't have this problem so they can be used by humans to get "human-like" positions without those problems, people that don't use engines this way wouldn't even be able to tell (some don't even play against other humans!)

[BrendanJNorman]

My opinion on this "human-like" debate has evolved over time...

Now I think the following:

To be human-like, a computer needs to make semi-regular mistakes.

Mistakes in:

- Strategy (allowing pawn structure to be destroyed, giving away bishop pair, losing control of only open-file, etc)

- Tactics (allow occasional "breakthroughs" when pieces are concentrated around the king...don't defend perfectly like an engine, allow even unsound sacs to occasionally "work" because it "missed something").

- Time management: (sometimes "think" for too long and get into "time trouble" and then replicate the resulting increased chance for blunders due to this)

All the while, this mistake-making "human-like" engine needs to have enough knowledge to play the human "patterns" that appear so often such as pawn storms, exchange sacrifices, pawn sac on e6/d6 to slow development etc.

All of this above needs to happen while maintaining a consistently HIGH average centipawn loss per move.

Because when it starts to play too strongly, it is no longer playing human-like, it is engine-like.

Conversely, when humans start to play with too *few* errors, we can call them engine-like as well.

I am well over 2400 on liChess (which is higher than top 1% in terms of rating) and am still absolutely feeble against most engines someone would call "human-like".

To play human-like chess is to play weak chess (compared to other engines) - weak chess that is full of recognizable patterns, and with super low calculation depth (1-2 moves in most positions).

A determined enough developer could create a *real* human-like engine, but it requires a fundamental paradigm shift and to absolutely abandon Elo as measured against computers.

The focus needs to be on reproducing human *mistakes* and in working from there backwards.

Just my 2 cents.

[Frank Quisinsky]

Hi there,

same opinion as Fritz 0!
The style of engine was not changed with a NN file.
Easy to see in stats!

If the style is changed with NN (newer version):
The probability is high that the program lost his good known own face if the programmer made to many other changes.
Nothing to do with NN.

Human-like:

1. In my opinion an engine play "Human-like" if most complicated playing phases for best players (transition to the endgame) the knowledge is not perfect. A big problem for strongest engines because exactly here the strongest engines produced the bigger advantages to most of others engines.

2. In my opinion if an engine is strong with many pieces on board. The advantage from humans in strategy are all the strategy ideas after and during the first playing phase ... the openings!

If such programs goes back to strongest GM level (a good example is reduce Elo with nps) a perfect simulation of human-like can produced. Human like engines are: Fizbo, Spark, Wasp or Hakka. Speculative engines produced human-like also. Examples are Smarthink, Hiarcs or Fritz (Gingko).

Best
Frank

[Daniel Shawul]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

[Fritz 0]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

How so?

[dkappe]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

How so?

I’ll be releasing a new branch of a0lite soon that can use a UCI engine under multipv to produce value and policy (just like a leela NN). At depth 7 for the AB engine (cfish in this experiment), the play looks more like an mcts/nn engine than an AB engine.

This is all subjective, but my impression has been that the search — uct vs ab — has more to do with the playing style than the net.

[Daniel Shawul]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

How so?

Because neural networks are inspired by the biological neurons of the human brain. CNNs breakthrough (local connectivity) is inspired by what goes on the eye etc -- so many new NN architectures inspired by the human brain since then!

Humans are weak at look-ahed search (the best GM could probably calculate less than 10 positions per second) so if you use more of brute force search as in alpha-beta, you stray away more from being human-like.

Talkchess likes to debate what is not debatable, but it is clear to me big NN are more human like.

[Joerg Oster]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

How so?

I’ll be releasing a new branch of a0lite soon that can use a UCI engine under multipv to produce value and policy (just like a leela NN). At depth 7 for the AB engine (cfish in this experiment), the play looks more like an mcts/nn engine than an AB engine.

This is all subjective, but my impression has been that the search — uct vs ab — has more to do with the playing style than the net.

That's also my impression.

[Uri Blass]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

I do not feel that
neural networks imitate what is going in the brain.

I tried dragon elo levels when dragon elo level is based on NN and got stupid moves in handicap games.
No human with elo 1600 is going to lose with that huge handicap even against god.

I find normal engines with no NN more human like because I believe that with time control even with small number of nodes per move (small enough to have level of 1600 elo) they will practically not lose with queen and 2 rooks handicap.

[pgn][Event "Computer event"]
[Site "Somewhere on Earth"]
[Date "2022.04.25"]
[Round "1"]
[White "1600 elo Dragon 2.6.1 by Komodo Chess 64-bit "]
[Black "Dragon 2.6.1 by Komodo Chess 64-bit contempt100"]
[Result "0-1"]
[TimeControl "depth: 24"]
[Time "12:51:16"]
[Board "1"]
[Termination "mate"]
[FEN "1nb1kbn1/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]
[SetUp "1"]

1. g4 {+22.29/1 828 20} d6 {+18.55/21 19444 30768317} 2. e4 {+22.30/1 10 23} Nf6 {+18.72/21 16842 27012437}
3. f3 {+22.13/1 10 40} a6 {+18.75/21 19249 30103897} 4. d4 {+22.15/1 10 28} Nfd7 {+18.78/22 33366 52171961}
5. c4 {+22.22/1 10 35} b6 {+18.91/21 27577 42208245} 6. b4 {+22.23/1 10 34} Bb7 {+18.61/22 17148 25146867}
7. a4 {+22.19/1 10 35} e6 {+18.53/22 26727 40622079} 8. g5 {+22.19/1 10 35} d5 {+17.79/24 11213 15614277}
9. Kd2 {+14.23/1 10 351} dxe4 {+15.15/22 7969 10494102} 10. h4 {+18.17/1 10 154} Bxb4+ {+14.55/21 13507 17199385}
11. Ke3 {+18.17/1 10 5} exf3 {+14.49/22 7704 10086668} 12. h5 {+18.17/1 10 32} c5 {+13.76/22 15134 19370092}
13. Nxf3 {+14.86/1 10 182} cxd4+ {+12.57/19 4186 5389959} 14. Kxd4 {+13.83/1 10 11} h6 {+12.92/21 6878 8882942}
15. g6 {+14.93/1 10 123} Nc6+ {+12.30/20 6539 8703618} 16. Ke3 {+12.35/1 10 95} Bc5+ {+12.50/21 8275 10732025}
17. Kf4 {+12.19/1 10 46} fxg6 {+12.46/21 3909 5243716} 18. Be3 {+12.40/1 10 76} g5+ {+9.08/21 7965 9631959}
19. Kg3 {+9.53/1 10 103} Bxe3 {+9.69/22 3601 4287368} 20. Kh3 {+11.91/1 10 54} e5 {+8.70/23 28384 35232597}
21. Kg3 {+11.99/1 10 40} Nf6 {+8.32/22 15828 19925824} 22. Kg2 {+11.79/1 10 45} g4 {+7.23/23 8840 10921115}
23. Kg3 {+9.33/1 10 345} Ne4+ {+5.40/26 23958 29601493} 24. Kh4 {+9.20/1 10 25} Bf2+ {M+3/30 710 1698629}
25. Kxg4 {+6.39/1 10 3} Bc8+ {M+2/35 669 1559365} 26. Qd7+ {M-1/1 10 2} Bxd7# {M+1/99 19 297}
0-1

[/pgn]

[Fritz 0]

How else do explain 1-node big NN is IM to GM level?
Morevoer, neural networks imitate what is going on in the brain, so it is TRUE that they are more human like.
AB engines with tiny NNUE add a deep search so they are less human like compared to big NNs.

How so?

Because neural networks are inspired by the biological neurons of the human brain. CNNs breakthrough (local connectivity) is inspired by what goes on the eye etc -- so many new NN architectures inspired by the human brain since then!

Humans are weak at look-ahed search (the best GM could probably calculate less than 10 positions per second) so if you use more of brute force search as in alpha-beta, you stray away more from being human-like.

Talkchess likes to debate what is not debatable, but it is clear to me big NN are more human like.

Then I must be blind. The other day I played over the games of the Fritz 5.32 vs. Judit Polgar rapid match from 1999. I couldn't tell any significant difference in the playing style between a 20+ years old engine and a top human GM, except that the engine was obviously stronger tactically. If the engines from the 1990s had a normal-looking style (at least to me), what about modern, positionally sophisticated engines like Komodo?

Also, when I analyze certain positions with Stockfish (a NN engine), I often don't understand its choices and evaluations. Ok, the reason can be that I don't have enough chess knowledge to get it. But when I give the same position to Komodo 14, all suggestions and evals suddenly seem logical and comprehensible to me, especially if the search depth is not too high.