Thoughts on Komodo 11, Free engines and the future

[cdani]

why do I then see Andscacs 0.86 pick the French on 1.e4, and then, after 2.d4 d5, wants to change ed5 ed5?

Because those changes are hints, not absolute rules, so the other big amount of parameters also count. Anyway as told I should tune better this parameter, so who knows... Also if I try to analyze well what happens, sure there are other related stuff to improve to achieve the whole effect that you propose.

[Lyudmil Tsvetkov]

Because those changes are hints, not absolute rules, so the other big amount of parameters also count.

precisely, when programmers test some evaluation parameter, and get +2-3 elo from it, it might hide another 30 or so elo, it is just that the rest of the parameters are not quite passing.

still, it is a bit disturbing that SF and other top engines of nowadays would prefer lines top human players used back in the 1890s...

obviously, top engines are still pretty much in their Romantic stage.

I also guess that tuning at 2 hours instead of 1 min. per game would produce much better results. How much computer power would running a test at 2 hours per game with 50 000 games take?

[Lyudmil Tsvetkov]

[d]rnb2rk1/pp2nppp/4p3/q1ppP3/3P1P2/P1P5/2PB2PP/R2QKBNR w KQ - 0 9

and in this Winaver position, Andscacs seemingly performs better than SF and Komodo, which largely think it is close to 0.0.
Andscacs gives some 20-30cps white edge, but actually white is mostly winning.

the e5 feature is still there, but there is also a pointed white chain c3-d4-e5, with e5 pawn redundant again.

how do we tune the pointed chain with the e5 feature, when the e5 feature is part of both configurations?

I guess anything else apart from:
- if pointed chain, do not score e5 feature, will fail

otherwise, bonus becomes too big and test does not succeed.

how many good positional features have failed tests because of inherent redundancies?

why would Andscacs see on above position white advantage, while not choose French lines with an e5 push?

[corres]

[quote="Lyudmil Tsvetkov"]

well, I can not agree with even a single of your conclusions.
they might be learned, somehow agree with current (book) theory, but are completely untrue....

[/quote]

In my database I found 2744 games with that position.
From these games are 1376 ended 1:0, 716 ended 1/1, 652 ended 0:1 so
White advantage is obvious.
I think that the main problem is all chess engines are the result of "endogamy". The most of engines are tested against its predecessor and structure of programs have the same frame. Moreover humans are not a real enemy against engines so there is nothing to press authors of engines to make real improvements in evaluation.

[cdani]

I think that the main problem is all chess engines are the result of "endogamy". The most of engines are tested against its predecessor and structure of programs have the same frame. Moreover humans are not a real enemy against engines so there is nothing to press authors of engines to make real improvements in evaluation.

+10

[Lyudmil Tsvetkov]

I think that the main problem is all chess engines are the result of "endogamy". The most of engines are tested against its predecessor and structure of programs have the same frame. Moreover humans are not a real enemy against engines so there is nothing to press authors of engines to make real improvements in evaluation.

+10

another approach is: kingside fianchetto, pointed chains, doubled pawns, those are really just some stupid concepts, most likely untrue and not applicable at all in chess engines.
engines know better

and truly, in my engine doubled pawns simply do not work...

[Lyudmil Tsvetkov]

well, I can not agree with even a single of your conclusions.
they might be learned, somehow agree with current (book) theory, but are completely untrue....

In my database I found 2744 games with that position.
From these games are 1376 ended 1:0, 716 ended 1/1, 652 ended 0:1 so
White advantage is obvious.
I think that the main problem is all chess engines are the result of "endogamy". The most of engines are tested against its predecessor and structure of programs have the same frame. Moreover humans are not a real enemy against engines so there is nothing to press authors of engines to make real improvements in evaluation.

thanks for the feedback, robert.

[d]6k1/5pp1/4p3/4P1P1/8/8/8/6K1 w - - 0 1

and I suppose, with this setup, white won all.

[Lyudmil Tsvetkov]

[d]6k1/8/8/8/4p3/4P3/5PP1/6K1 w - - 0 1

I suppose black, in spite of tempo down, will also have advantage in stats with the mirrored position.

[d]rnbqrbk1/ppp2pp1/3p1n1p/3P4/2P1p3/1PN1P2P/PBN1BPP1/R2Q1RK1 w - - 0 4

this might arise from the English, if white plays theoretically with e3 instead of the much better e4.

SF and Andscacs give here 60cps! white advantage, Komodo only 40cps, but, as a matter of fact, white is already lost.

I guess testing this position with thousands of games will not be productive, as you need some 20 excellent moves, and not random ones with 1 sec per move, until black starts winning.

the very best of modern chess software: giving 60cps advantage for the side that trails behind with the same score.

[Lyudmil Tsvetkov]

indeed, a short shootout with Komodo 8 brought 5 white wins, 3 draws and 2 black wins.

most of the games Komodo chose Nd4 c5? dc6 Nc6, after which the game opens and white retains the edge.

still, the only game that Komodo chose to play only on the king side, the right approach, with Nd4 g6, followed by Bg7, etc., black won quite convincingly.

[pgn][Event "Shootout (Komodo864-bit, Blitz 1m) "]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "Andscacs 0.86134"]
[Black "owner"]
[Result "0-1"]
[SetUp "1"]
[FEN "rnbqrbk1/ppp2pp1/3p1n1p/3P4/2P1p3/1PN1P2P/PBN1BPP1/R2Q1RK1 w - - 0 4"]
[PlyCount "98"]

4. Nd4 {0.64/15 2} g6 {0.55/16 2} 5. Qd2 {0.56/16 1 (Qc2)} Nbd7 {0.53/17 3
(Bg7)} 6. Bd1 {0.56/17 2} a5 {0.53/18 2 (Bg7)} 7. Bc2 {0.51/17 1} Bg7 {0.55/17
1} 8. a3 {0.55/17 3 (Rad1)} Ra6 {0.56/17 2} 9. Rfd1 {0.54/16 1} Ra8 {0.54/17 4}
10. Qe2 {0.55/17 3} Nb6 {0.52/17 1} 11. b4 {0.47/17 4} Re5 {0.38/14 1 (Bd7)}
12. Ncb5 {0.36/14 2 (Ndb5)} Rg5 {0.30/15 1} 13. Kh2 {0.14/14 1 (Kh1)} Bd7 {0.
05/14 1} 14. Bb3 {-0.01/15 3 (bxa5)} a4 {0.08/16 3} 15. Ba2 {-0.09/15 0} Ng4+ {
0.01/14 2 (Rh5)} 16. hxg4 {-0.14/16 2 (Kh1)} Bxg4 {0.13/16 1} 17. Qe1 {-0.13/
19 1} Bxd1 {0.13/16 1 (Rh5+)} 18. Qxd1 {-0.12/17 1 (Rxd1)} Rh5+ {-0.24/15 1}
19. Kg1 {-0.24/19 3} c5 {-0.28/17 3 (Qh4)} 20. dxc6 {-0.12/17 0} bxc6 {0.12/16
1} 21. Nc3 {-0.12/18 1} Qh4 {0.12/17 1} 22. f4 {-0.12/16 1} c5 {0.12/17 3} 23.
bxc5 {-0.12/18 1} dxc5 {0.12/17 1 (Qh1+)} 24. Nde2 {-0.12/17 0} Nc8 {-0.12/17
3 (Rb8)} 25. Qc2 {-0.12/18 1} Rb8 {-0.01/17 1} 26. Rd1 {-0.12/18 1} Ne7 {0.03/
16 1} 27. Bc1 {-0.12/18 1} Nf5 {-0.14/16 1} 28. Nd5 {-0.39/16 1} Kh7 {-0.69/16
0} 29. Rf1 {-0.63/16 1} Qg4 {-0.90/16 1} 30. Nec3 {-0.96/17 2} Ng3 {-0.61/14 1}
31. Re1 {-0.56/16 0} g5 {-0.91/17 2} 32. Qd1 {-1.03/16 1} Qh4 {-1.19/17 1} 33.
Qc2 {-1.58/17 2} Nf5 {-1.59/15 0} 34. Bd2 {-1.44/17 2} Qh2+ {-1.63/18 3} 35.
Kf1 {-1.83/16 1} Nh4 {-1.77/15 0} 36. Ke2 {-1.37/15 2} Qxg2+ {-2.02/14 0 (f5)}
37. Kd1 {-2.17/15 0} f5 {-2.29/16 1} 38. Re2 {-2.77/15 1} Qh1+ {-2.78/15 0} 39.
Re1 {-3.09/15 1} Nf3 {-3.09/16 0} 40. Rxh1 {-3.26/15 0} Rxh1+ {-3.29/16 0} 41.
Ke2 {-3.29/17 0} Rh2+ {-3.53/17 1} 42. Kd1 {-3.85/15 0} Rxd2+ {-3.77/16 0} 43.
Qxd2 {-3.94/17 0} Nxd2 {-4.03/16 0} 44. Kxd2 {-4.05/17 0} Bxc3+ {-4.29/18 1}
45. Nxc3 {-4.43/16 0} Rd8+ {-4.38/15 0} 46. Kc1 {-4.55/16 1 (Ke2)} Rd3 {-4.55/
16 0} 47. Nxa4 {-4.60/16 0} Rxa3 {-4.69/16 0} 48. Nxc5 {-4.81/18 1} g4 {-5.75/
16 1 (Rxa2)} 49. Bb1 {-6.04/15 0} g3 {-6.35/15 0} 50. Nxe4 {-6.35/16 0} g2 {-6.
41/15 0} 51. Kb2 {-6.84/15 1} g1=Q {-6.89/16 0} 52. Ng5+ {-6.96/16 0} Kg7 {-7.
08/17 0} 0-1

[/pgn]

difficult to tune engines, when they choose the wrong plan of play...

[cdani]

I will review all this, but this requires time
You can take the latest Andscacs from http://www.andscacs.com/en/en_index.html Is clearly stronger.
Thanks!