Glaurung 0.1.5 UCI and resign by Polyglot

[Igor Gorelikov]

Glaurung 0.1.5 UCI and resign by Polyglot

In the running Final qualifier for New IL-7 I've got 4 games with strange
results. Glaurung 0.1.5 UCI resigns in equal positions.

Actually it is done by Polyglot. Of course, the resign function of Polyglot
is far from perfection. It is based on one move score which can be
dangerous. But why Glaurung shows losing score?

Here is one of the game.

[Event "New IL-7, Final Qualifier"]
[Site "P3 1GHz 256MB"]
[Date "2004.12.07"]
[Round "7.1"]
[White "Glaurung 0.1.5 UCI"]
[Black "Snitch 1.0.8"]
[Result "0-1"]
[TimeControl "1200+3"]
[PairNo ""]

1.Nc3 Nf6 2.d4 d5 3.Bf4 e6 4.e3 Bb4 5.Bd3 Bxc3+ 6.bxc3 Nc6 7.Nf3 O-O 8.O-O
b6 9.Ne5 Nxe5 10.Bxe5 Nd7 11.Bf4 Bb7 12.Qf3 f5 13.Qg3 Rc8 14.c4 Nf6 15.Be5
Ng4 16.cxd5 Nxe5 17.Qxe5 Qxd5 18.Qxd5 {White resigns} 0-1

[diag]2r2rk1/pbp3pp/1p2p3/3qQp2/3P4/3BP3/P1P2PPP/R4RK1 w - - 0 18[/diag]

Here is a cut from Winbord debug:

Code: Select all

13300 <first : 10 -3 159 163009 Qxd5 Bxd5 f3 Rcd8 Kf2 Kf7 Kg3 Kf6 Kf4 g5+ Kg3
14170 <first : 10 -9998 252 260731 Qxd5 Bxd5 f3 Rcd8 Kf2 Kf7 Kg3 Kf6 Kf4 g5+ Kg3
14170 <first : move e5d5


Glaurung shows the losing by mate score (-9998) in the equal position! Why?

Of course, the resign option is switch off now and all the 4 games will be
replayed.

====================================================
For games and more info on the Infinite Loop apply to
http://freehosting.hostrave.com/p/iggor/home.php

Igor

[Guenther Simon]

Hi Igor,

This is known and I posted it already here
and at my site (test tourneys).
BTW the same can also happen for Adam, no idea though
if this still applies for the newest version.
(I would disable resign for Adam to be safe)

Best regards,
Guenther

[Igor Gorelikov]

Hi Guenther ,

Thanks for info.
I checked all the games of Adam 2.6b UCI and don't find a wrong resign.

Igor

[Laurens Winkelhagen]

My own engine janwillem showed (until version 1.08) a similar behavior. In his case it came about in the following manner:

Janwillem precalcuates a time to think about a position. When (at a certain amount of nodes and at a certain depth) janwillem discovers that his time for this move is up, it switches a variable and returns to the previous depth. Where janwillem would find again that the time is up and returns to an even swallower depth. I'm pretty sure there are more engines that work this way. but now the problem:

Jan willem would return alpha in such cases. and alpha may be a pretty low score, for instance when it hasn't been corrected upwards at all:-) This alpha keeps getting returned to swallower depths and will eventually come to the root.

At that point Janwillem would output it's best move sofar and the accompanying return score, which could be -32000, if janwillem was at the beginning of a new depth for it's iterative search:-)

Of course I'm not sure, and I'm definately not a better coder than Tord, but this could also be the case for Glaurung.
I hope it helps:-)

Laurens.

[Uri Blass]

I have also a flag(stop_search) that tells me to stop search but if that flag is up(stop_search!=0) I do the following steps:

1)Undo the move
2)Returning 0 in case that I am not in the root position(it is practically not important what I return in that case and 3 is what is important).
3)Returning the previous score in case that I am in the root position and the last move that I searched cause change in the score(it is clear that I cannot trust change in the score in this case because I did not finish to search all the possibilities).