Gull 3 Linux+Syzygy and Fathom released.

[carldaman]

yes, I know this seemingly immanet syzygy behaviour ... but this is not the case
at least with a mingw 5.3 compile under Windows

the position is mate in #7
e.g. 1.Nc5 / 1.Nc4 / 1. Nf4 / Ba4

no other engine with syzygy 5- or 6-men plays 1.Re7


with this Gull selfplay goes

1.Re7 #(#0) Kc8 # (#0)
2.Rc7+ #(#0) Kb8 # (#0)
3. Kc1 #(#0) Ka8 # (#0)
and so on ...
evaluation stays always at #0 and not thinking (Fritz15 gui)

in Arena 3.51 or cli it's different but confused too

switching the hash to other sizes like 64, 128, 1024 leads to other evaluations or even exceptions of the engine

think there is a hidden bug, maybe only in windows

+1

[basil00]

yes, I know this seemingly immanet syzygy behaviour ... but this is not the case
at least with a mingw 5.3 compile under Windows

the position is mate in #7
e.g. 1.Nc5 / 1.Nc4 / 1. Nf4 / Ba4

Mate-in-7 is 14plys, whereas white can force a zeroing move in 7plys (i.e. DTZ=7) by playing 1.Re7. Essentially white's "plan" is to force black to capture the rook and then win the KNBvK endgame.

Note that 1. Re7 makes sense if the 50-move rule is about to come into effect. If white plays the Nalimov moves then it will be a draw.

Here is the expected self-play (Fathom output):

[pgn][Event ""]
[Site ""]
[Date "??"]
[Round "-"]
[White "Syzygy"]
[Black "Syzygy"]
[Result "1-0"]
[FEN "3k4/8/8/8/8/3N4/2BKR3/8 w - - 92 1"]
[WDL "Win"]
[DTZ "7"]
[WinningMoves "Re7, Ba4"]
[DrawingMoves "Kc1, Kd1, Ke1, Kc3, Ke3, Re1, Rf2, Rg2, Rh2, Re3, Re4, Re5, Re6, Re8, Bb1, Bd1, Bb3, Nc1, Ne1, Nb2, Nf2, Nb4, Nf4, Nc5, Ne5"]
[LosingMoves ""]

1. Re7 Kc8 2. Rc7+ Kb8 3. Kc1 Ka8 4. Rb7 Kxb7 5. Kd2 Kc6 6. Ke3 Kd6 7. Ke4 Ke6 8. Bb3+ Kf6 9. Nf4 Kg7 10. Ke5 Kh7 11. Kf6 Kh6 12. Ba2 Kh7 13. Ng6 Kh6 14. Bg8 Kh5 15. Ne5 Kh4 16. Kf5 Kg3 17. Ng4 Kg2 18. Bc4 Kf3 19. Bf1 Kg3 20. Be2 Kh4 21. Nf2 Kg3 22. Nd1 Kh4 23. Kf4 Kh3 24. Bf3 Kh2 25. Nf2 Kg1 26. Kg3 Kf1 27. Nd3 Kg1 28. Be2 Kh1 29. Nf2+ Kg1 30. Nh3+ Kh1 31. Bf3# 1-0[/pgn]

no other engine with syzygy 5- or 6-men plays 1.Re7

Other engines use search and then syzygy to filter the engine moves. This results in more "natural" play. This is not what Gull does though.

with this Gull selfplay goes

1.Re7 #(#0) Kc8 # (#0)
2.Rc7+ #(#0) Kb8 # (#0)
3. Kc1 #(#0) Ka8 # (#0)
and so on ...
evaluation stays always at #0 and not thinking (Fritz15 gui)

OK the moves are correct but the eval is wrong? I assume (#0) is the draw score?

I noticed a similar problem with xboard and this patch fixed the problem:
https://github.com/basil00/Gull/commit/ ... c34fa6cdae
If the above patch does not work then it is a new bug.

[basil00]

Why not run a search before, instead of auto-playing the TB move? Other top engines do that and achieve more realistic EG play.

This would be ideal. But it is not simple to implement with Gull's opaque source code.

[basil00]

Ran 2 gauntlets with Gull and Gullzygy.
40/1 tc
+11 elo

Thanks!

[stuwph]

no other engine with syzygy 5- or 6-men plays 1.Re7

Other engines use search and then syzygy to filter the engine moves. This results in more "natural" play. This is not what Gull does though.

I see, important to know

OK the moves are correct but the eval is wrong? I assume (#0) is the draw score?

I noticed a similar problem with xboard and this patch fixed the problem:
https://github.com/basil00/Gull/commit/ ... c34fa6cdae
If the above patch does not work then it is a new bug.

the patch was implemented already

maybe I am wrong but I still think something is not working correctly
will research some more on the weekend

[Dan Cooper]

From TCEC chat

Anton Mihailov (Tournament Director)

Yes, I am considering it. The main problem is that there is no testing before stage 2. So if the (minor) change in Gull has problems, it can ruin the whole event for the engine. So I guess another poll on facebook with brainstorning pluses and minuses can be handy

[Dan Cooper]

From TCEC chat

Anton Mihailov (Tournament Director)

Yes, I am considering it. The main problem is that there is no testing before stage 2. So if the (minor) change in Gull has problems, it can ruin the whole event for the engine. So I guess another poll on facebook with brainstorning pluses and minuses can be handy

My response to this would be that every engine gets to update between rounds, so Gull should be no exception.

[basil00]

Anton Mihailov (Tournament Director): So I guess another poll on facebook with brainstorning pluses and minuses can be handy

This is a reasonable idea. I plan to have a stable release (a stage 2 candidate) within the next week or so.

[basil00]

A new optimized build for Linux has been tested, see here:
https://github.com/basil00/Gull/commit/ ... ffa9ff10d8

Code: Select all

LTC (60"+0.6", 1-thread, 6-man syzygy):

    Score: +1728-1351=6921
    Number of games: 10000
    Winning fraction: 0.51885
    Elo difference: +13.1045
    LOS: 1

STC (10"+0.1", 1-thread, 6-man syzygy):

    Score: +2040-1791=6169
    Number of games: 10000
    Winning fraction: 0.51245
    Elo difference: +8.65293
    LOS: 0.999971

I will release binaries (Linux and Windows) as soon as I can find a suitable Windows machine to build. The intention is that this will be the "release candidate".

[Dan Cooper]

The exes for the next stage are likely due soon. I would email him the best BMI2 build you have and at least hope he makes a decision to include it.