playing like a human should do it

[MikeGL]

yes I agree, Bf8 is nothing and very simple for any engines because Rg8 move was already played on the board.
But that line (with Bf8) cannot be reached without first figuring out Rg8 prior to Bf8.
Only after Rg8 can Bf8 be very useful and logical later, which clears the g-file.
Again, Bf8 is very easy even for weak engines because 11... Rg8 is already waiting to be activated.
If there's one strong positional move made on this game, I think this is 11... Rg8! of black which is not easy to find for A/B engine and the strength of this Rg8 will be uncorked later after Bf8. In short, Bf8 is nothing and Rg8 is the real deal.

Beautiful game. Thanks for sharing.
It is a very deep tactic in my opinion starting with Bf8! provoking Nb5 -> Nc7 fork

A/B engines see Bf8 almost instantly, even without NNUE.

All I see in this game is a bad blunder (Nb5) by an old engine. That move is the most human-like of all ... a mistake.

[peter]

yes I agree, Bf8 is nothing and very simple for any engines because Rg8 move was already played on the board.
But that line (with Bf8) cannot be reached without first figuring out Rg8 prior to Bf8.
Only after Rg8 can Bf8 be very useful and logical later, which clears the g-file.
Again, Bf8 is very easy even for weak engines because 11... Rg8 is already waiting to be activated.
If there's one strong positional move made on this game, I think this is 11... Rg8! of black which is not easy to find for A/B engine and the strength of this Rg8 will be uncorked later after Bf8. In short, Bf8 is nothing and Rg8 is the real deal.

Cfish with Use NNUE Pure at a single try with empty hash (LC0 had a move history already in the game, even if it seems, it was only 4 moves long?)

[d]r3k2r/pp1nppb1/2pp1n1p/q7/P2PP1b1/2NQ1N2/1PP1BPP1/R1B2RK1 b kq - 0 1
Analysis by Cfish 040920 64 POPCNT NUMA:

Code: Select all

11...0-0-0 12.Ne1 Rdg8 13.Bxg4 Nxg4 
  =  (0.22)   Depth: 7/7   00:00:00  42kN
11...Rg8 12.Nh2 Bxe2 13.Nxe2 c5 14.Be3 cxd4 15.Nxd4 
  +/=  (0.62)   Depth: 8/10   00:00:00  96kN
...
11...Rg8 12.Rd1 Qh5 13.Bf4 Bxf3 14.Qxf3 Qxf3 15.Bxf3 h5 16.a5 a6 17.Ne2 e5 18.dxe5 dxe5 19.Bg5 Bf8 20.Bxf6 Nxf6 21.Ng3 Rd8 22.Rxd8+ Kxd8 23.Rd1+ Kc7 24.Nxh5 Nxh5 
  +/-  (1.13)   Depth: 24/33   00:00:12  112MN
11...Qh5 
  +/-  (1.05 ++)   Depth: 25/39   00:00:15  143MN
11...Rg8 
  +/-  (0.97 ++)   Depth: 25/39   00:00:20  184MN
11...Rg8 12.Rd1 
  +/-  (1.21 --)   Depth: 25/39   00:00:21  191MN
11...Qh5 
  +/-  (1.02 ++)   Depth: 25/39   00:00:22  201MN
11...Qh5 12.Rd1 e6 13.a5 Rg8 14.Bf4 Bf8 15.a6 b5 16.e5 dxe5 17.Nxe5 Nxe5 18.dxe5 Bxe2 19.Nxe2 Nd5 20.c4 bxc4 21.Qxc4 Rc8 22.Rd4 Rg4 23.f3 Rxf4 24.Nxf4 Qxe5 25.Nxd5 
  +/-  (0.86)   Depth: 25/39   00:00:26  238MN
11...Qh5 12.Rd1 
  +/-  (0.94 --)   Depth: 26/39   00:00:27  242MN
11...Qh5 12.Rd1 
  +/-  (1.02 --)   Depth: 26/39   00:00:27  244MN
11...Rg8 12.Rd1 e6 13.b4 Qh5 14.a5 a6 15.Bf4 Bf8 16.e5 dxe5 17.dxe5 Nd5 18.Nxd5 exd5 19.c4 dxc4 20.Qe3 Rg6 21.Rdc1 b5 22.axb6 Nxb6 23.Qxb6 Bxf3 24.Bxf3 Qxf3 25.Bg3 Qd5 26.Rd1 c5 27.Qc7 
  +/-  (1.10)   Depth: 26/39   00:00:28  256MN
...
11...Rg8 12.Rd1 Bf8 13.Nh2 Bh5 14.Rb1 Bxe2 15.Qxe2 Qh5 16.Qxh5 Nxh5 17.g3 e6 18.Nf3 Nhf6 19.Kg2 d5 20.e5 Ng4 21.Ne2 f6 22.Nf4 fxe5 23.dxe5 Ndxe5 24.Nxe5 Nxe5 25.Nxe6 Kd7 
  +/-  (1.24)   Depth: 28/42   00:00:51  458MN
...
11...Rg8 12.Rd1 e6 13.Bf4 Bf8 14.d5 cxd5 15.exd5 e5 16.Be3 Be7 17.Nh2 Bh5 18.Bxh5 Nxh5 19.Qe2 Ng7 20.Ng4 f5 21.b4 Qxb4 22.Nxh6 Rh8 23.Nb5 Qe4 24.Nc7+ Kf8 25.Ne6+ Nxe6 26.dxe6 Nf6 27.f3 Qh4 28.Nxf5 
  +/-  (1.06)   Depth: 29/46   00:00:59  532MN
11...Rg8 
  +/-  (0.98 ++)   Depth: 30/33   00:01:01  550MN
11...Qh5 
  +/-  (0.89 ++)   Depth: 30/41   00:01:18  703MN
...
11...Qh5 12.Rd1 Rg8 13.Bf4 Qg6 14.Nh4 Qh5 15.Nf5 Bf8 16.Bg3 Bxe2 17.Nxe2 Qg6 18.Nf4 Qg5 19.a5 e6 20.a6 b5 21.Nh4 Qg4 22.Qe3 Be7 23.Nf3 h5 24.b4 Nb6 25.d5 Nc4 26.Qd3 h4 
  +/-  (0.97)   Depth: 30/45   00:01:34  846MN
11...Qh5 12.Rd1 
  +/-  (1.05 --)   Depth: 31/46   00:01:40  893MN
11...Rg8 
  +/-  (0.97 ++)   Depth: 31/46   00:01:45  944MN
11...Rg8 12.Rd1 
  +/-  (1.13 --)   Depth: 31/46   00:01:51  994MN
11...Rg8 12.Rd1 Qh5 13.Bf4 Qg6 14.Nh4 Qh5 15.Nf5 Bf8 16.a5 a6 17.Bg3 Bxe2 18.Nxe2 Qg4 19.f3 Qg5 20.Ne3 e6 21.Kf2 d5 22.exd5 Nxd5 23.Nxd5 exd5 24.Qe3+ Qxe3+ 25.Kxe3 0-0-0 26.Kf2 Re8 27.Re1 Bb4 28.c3 Be7 29.Nf4 Nf8 30.Re3 
  +/-  (1.17)   Depth: 31/46   00:02:00  1073MN
...
11...Rg8 12.Rd1 Bf8 13.Nh2 h5 14.Bd2 Bxe2 15.Nxe2 Qa6 16.Qxa6 bxa6 17.f3 h4 18.b4 e6 19.Rdb1 c5 20.c3 h3 21.g4 Rc8 22.Kf2 cxd4 23.cxd4 Nb6 24.Nf1 Nc4 25.b5 Nxd2 26.Nxd2 
  +/-  (1.05)   Depth: 32/48   00:02:09  1150MN

And after switching to MultiPV=2, output in depth 34:

r3k2r/pp1nppb1/2pp1n1p/q7/P2PP1b1/2NQ1N2/1PP1BPP1/R1B2RK1 b kq - 0 1

Analysis by Cfish 040920 64 POPCNT NUMA:

1. +/- (1.34): 11...Rg8 12.Rd1 Qh5 13.Bf4 e5 14.Bg3 a5 15.Rab1 Bf8 16.b4 axb4 17.Rxb4 exd4 18.Nxd4 Bxe2 19.Ndxe2 Nc5 20.Qd4 Ng4 21.f3 Ne5 22.Qf2 Qg6 23.Bh2 Be7 24.Ng3 Bh4 25.Nce2 Rxa4 26.Rxa4 Nxa4 27.Qd4 Nc5 28.Kh1 Ned7 29.Qxd6 Qxd6 30.Rxd6 Bxg3 31.Nxg3 Rg6 32.Nf5 h5 33.Rd4 Ne6

2. +/- (1.13): 11...Qh5 12.Rd1 Rg8 13.Bf4 e5 14.Bg3 a5 15.dxe5 dxe5 16.Qd6 0-0-0 17.Nxe5 Nxe4 18.Bxg4 Nxd6 19.Bxh5 Nxe5 20.Rab1 Rge8 21.b3 Kc7 22.Ne2 Ng6 23.Rd3 Be5 24.f4 Bf6 25.Rbd1 Nf5 26.Bf2 Ng7 27.Bg4 h5 28.Rxd8 Rxd8 29.Rxd8 Bxd8 30.Bh3 b5 31.Be3 bxa4 32.bxa4 Ne8 33.Kf2 Nd6 34.g3 Bf6 35.Kf3 Ne7 36.Bg1 Nd5

...Rg8 and ...Qh5 might well transpose and anyhow White is still better at this point of the game, isn't it?

What I'd be interested in was mainly, how comes this strange opening line (3...h6?, 4.h3?...) ?
Engines start pondering at 7th. if I interpret time- and depth- commentaries stored in .pgn correctly, so I guess up to there it was a test- line given?

As Louis said already, most human-like moves are the blunders in this game one more time, as it is mostly the case, when engines play and humans do so too over the board.
I'd like to see Thorsten play at this TC against ancient Rybka, need not be version 4 at all, any earlier one would do.
Even if he was allowed to have White with this strange opening line and Rybka Black from move 7 onwards.
Or one of his beloved old dedicateds with porgrams full of knowledge and wisdom, that humans had told them back in glorious days of CC.

regards

[M ANSARI]

I really doubt any human could have played like that black side played. Humans simply cannot calculate things accurately when there is a combination of deep tactics and tempos to calculate simultaneously. Intuition might guide them to the correct path on some of the moves, and of course if there is home preparation they can jump into crazy complications, but one missed tempo and all is lost.

[mclane]

The AB engines in that tournament get 6 cores (!) and should not be „that“ weak at all.

I guess Lc0 will maybe win against them all. The NPS of Lc0 with this network dropped from 30.000 (LC0 0.26.0 with 703350) to 3.000.
So I guess the network is much bigger then the 703350 I used before.

What my point is that the games of LC0 remind me on a human player.
LC0 is not playing boring machine chess.
I can maybe come with more samples .

[mclane]

Another example.

This time an 8 bit engine by Ed Schroeder.
It got a little turbo by running it from 5 (original speed) mhz to 100 MHz with a new turbo kit.

[pgn]
[Event "40/120 100 mhz"]
[Site "SCW"]
[Date "2020.09.04"]
[Round "17"]
[White "Mephisto Polgar sel3 100 mhz"]
[Black "Novag Star Diamond "]
[ECO "D25"]
[Result "1-0"]

1. d4 d5 2. c4 dxc4 3. Nf3 Nf6 4. e3 Bg4 5. Bxc4 e6 6. h3
Bh5 7. O-O Nbd7 8. Nc3 Bd6 9. e4 e5 10. Be2 O-O 11. dxe5
Nxe5 12. Nd4 Bxe2 13. Qxe2 {"} Bc5 14. Nb3 Bb4 15. Bg5 Bxc3
16. bxc3 Qd6 17. Rad1 Qc6 18. Rd4 Rfd8 19. f3 Qb6 20. Be3
Qe6 21. Nc5 Qc6 22. Rfd1 Rd6 23. Nd3 Nfd7 24. Nb2 Nb6
25. f4 Nec4 26. Nxc4 Nxc4 27. Rxd6 cxd6 28. Bd4 Re8 29. Qg4
g6 30. Qh4 Re6 31. f5 Rxe4 32. Qh6 f6 33. fxg6 Qd7 34. Bxf6
a5 35. Qg5 hxg6 36. Qxg6+ Kf8 37. Rf1 Qf7 38. Bg7+ {Matt in
6} Ke7 39. Rxf7+ Kd8 40. Qxe4 a4 41. Bf6+ Kc8 42. Qxb7# 1-0
[/pgn]


The NOVAG star diamond has an Elo rating of 2186.

It runs a H8s chip with 16 bit on 25 MHz with 118 kB hash tables, engine by David Kittinger.

I did especially like the king attack the Schroeder engine creates here.

The original ssdf ELO rating of the 5 MHz Polgar is 1972 ELO and it runs a 6502 8 bit cpu without hash or any other tricks.
Of course the turbokit that speeds the cpu from 5 to 100 MHz is good for arround 2 plys deeper search in general.

[peter]

What my point is that the games of LC0 remind me on a human player.
LC0 is not playing boring machine chess.

As for some moves and games you're right of course, but as for some others it's just the other way round.

I support your aim, but I guess human guidance of net- development was the way to go at the moment primarily.

With more human intervention as for guided net- training instead of letting nets get bigger and bigger by selflearning and reinforcement only LC0- wise ("Zero- Gospel", simply disgusting in the meantime, don't you think so too, Thorsten? )
On the other hand more and more selflearning and reinforcement learning only of NNUE nets is even after such a short time of NNUE- development already getting saturating as it seems.
Adding .pgn- learning, which would have much more potential to me, would already be the way to go for humans now, as for my pov.

http://talkchess.com/forum3/viewtopic.p ... 53#p859353

That's a point for Thorsten Czub btw., regarding the other thread of his about human and engine- playing, programming and selflearning.
At least testing should be somewhat more human- guided and less automated and selfplay- guided, not to speak about guidance of net- training, which nowadays might have most potential, if it isn't restricted to reinforcement and selflearning only, but starts having steps into .pgn- learning also

Even if of course there's much room for human intervention in coding too, especially with better and better eval, hard- coded and or net-supported, helping to refine the search also.

Selfquoting by humans as well as selflearning by humans and engines is simply ok now and then too to me.

regards

[zullil]

Yes, the machine who taught itself how to do chess plays more human then any human effort to produce a chess engine,

This is indeed disapointing and shows the situation we have,

Well, who taught the machine how to teach itself to play chess?

I think that was Skynet by Cyberdyne Systems.

[Frank Quisinsky]

Hi,

the human like style is very easy to explain.

Human's have strengths with many pieces on board.
On schedule ideas, positional or strategic ideas.

With lesser pieces on board best chess programs are 500-600 Elo stronger, transposition into endgame. Interesting for analyzing games but not for self playing.

If an engine will produced a human-like-style the endgame should be have around the same strength as the middlegame. In earlier middlegame king safty should be good and aggressive moves an engine should play with many pieces on board. That's what human's like to see and to play.

That's the human style best players the World have.

LCO have nothing from it.
Playing strength in endgame / transposition into endgame is 500-600 over best chess players.
Many tactical holes in complicated opening systems, likes dutch or king indian others.

LCO is strong with pawn structures, in closed positions and transposition into endgame.
Stronger with lesser pieces on board.

TOP-Programs produced the human-style:
Hiarcs, Pedone, Chiron, Wasp, Arasan ... Andscacs is in endgames 150 Elo to strong.

And here are important that this group of engines can be set from 1500 - 2800 Elo on DGT-Pi or for chess computer developments. Because engines, produced a human-like-style are most interesting for chess computers and self playing.

LCO have nothing from it!
Can be very easy see in good statistics!

Best
Frank

[pohl4711]

Yes, the machine who taught itself how to do chess plays more human then any human effort to produce a chess engine,

This is indeed disapointing and shows the situation we have,

The normal AB engines create ELO, but they play a machine chess.

The only engine that plays human alike chess on this level is in the moment LC0.

Agreed. What is especially cool is, that Lc0 plays incredible strong (using a big 30x384 net), with calculating the root-node (board position), only, without any search. I would estimate the strength around 2200-2300 Elo. Without any search. Nothing could be more "non-computer"- and human-like, than this! That is so awesome. And so fun to play against. In your world, it means: What Mephisto 3 promised, Lc0 made come true.

[mclane]

here another interesting game of then polgar against the Novag Diablo 68000:

[pgn]
[Event "40/120 100 mhz"]
[Site "SCW"]
[Date "2020.09.06"]
[Round "18"]
[White "Mephisto Polgar sel3 100 mhz"]
[Black "Novag Diablo 68000 sel4"]
[ECO "A66"]
[Result "1-0"]

1. d4 Nf6 2. c4 c5 3. d5 e6 4. Nc3 exd5 5. cxd5 d6 6. e4 g6
7. f4 Bg7 8. e5 dxe5 {"} 9. fxe5 Nfd7 10. e6 fxe6 11. dxe6
Qh4+ 12. g3 Bxc3+ 13. bxc3 Qe7 14. Bh3 Nf8 15. Nf3 Nxe6
16. O-O O-O 17. Bh6 Re8 18. Qb3 Nc6 19. Rae1 Kh8 20. Ng5
Ncd8 21. Qd5 Qc7 22. Rf6 Qe7 23. Rexe6 Bxe6 24. Qe5 {Matt
in 6} 1-0
[/pgn]