GullChess 2.1

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Norbert Raimund Leisner
Posts: 1445
Joined: Tue May 20, 2008 2:57 pm
Location: Augsburg - Germany

GullChess 2.1

Post by Norbert Raimund Leisner » Tue Jun 18, 2013 7:28 am

Hello,

GullChess 2.1 is out -
http://www.sdchess.ru/download_engines.htm

Here are the changes in reference to former versions:

Gull 2.1 @ 17.06.2013
Evaluation weights are optimized with the use of automated tuning (source code included).
Gull's evaluation is no longer almost identical to that of Ivanhoe.
Minor search & time management & SMP efficiency enhancements.

On SMP & Hyper-Threading:
By default Gull limits the number of processes to the number of phisical CPU cores availible.
This is done in order to avoid ELO drop due to significant SMP search overhead at bullet time controls.
However for long games & analysis it is recommended to turn HT on and manually set the number of processes (UCI option "Threads")
to equal the number of logical CPUs.

On the automated tuning (technical):
Tuning is done in two steps:
1. an estimate of the gradient is calculated;
2. line search is performed.
SPSA (just move a bit along the gradient direction instead of performing a line search) may also be quite effective,
though it's difficult to pick good parameters.
Fairly simple exit criteria are used for the line search. I don't employ SPRT because it's guaranteed to be optimal only for
elo=elo1 vs elo=elo0 hypothesis, whereas in this case we are interested in elo>elo0 vs elo<=elo0 test. However it's still possible
for SPRT (or some other test) to be better for some elo-elo0 values. Shouldn't be very difficult to run a numerical comparision
between various tests.


Best wishes,
Norbert

Norbert Raimund Leisner
Posts: 1445
Joined: Tue May 20, 2008 2:57 pm
Location: Augsburg - Germany

Re: GullChess 2.1

Post by Norbert Raimund Leisner » Tue Jun 18, 2013 7:37 am

but Gull 2.1 exe does not run on Windows XP 3 (32-bit)

error message (translated from German) :
Gull 2.1 exe is not an allowed win32-application

User avatar
velmarin
Posts: 1579
Joined: Mon Feb 21, 2011 8:48 am

Re: GullChess 2.1

Post by velmarin » Tue Jun 18, 2013 7:47 am

Norbert Raimund Leisner wrote:but Gull 2.1 exe does not run on Windows XP 3 (32-bit)

error message (translated from German) :
Gull 2.1 exe is not an allowed win32-application

Probe now, ThinkingALot in Inmortal Forum:

Code: Select all

Re&#58; Gull
I recompiled the 32-bit version so it doesn't require any additional DLLs now. Gull 2.1.zip file on sourceforge is updated.
Last edited by ThinkingALot; Today at 07&#58;54.

User avatar
geots
Posts: 4790
Joined: Fri Mar 10, 2006 11:42 pm

Re: GullChess 2.1

Post by geots » Tue Jun 18, 2013 8:03 am

Norbert Raimund Leisner wrote:Hello,

GullChess 2.1 is out -
http://www.sdchess.ru/download_engines.htm

Here are the changes in reference to former versions:

Gull 2.1 @ 17.06.2013
Evaluation weights are optimized with the use of automated tuning (source code included).
Gull's evaluation is no longer almost identical to that of Ivanhoe.
Minor search & time management & SMP efficiency enhancements.

On SMP & Hyper-Threading:
By default Gull limits the number of processes to the number of phisical CPU cores availible.
This is done in order to avoid ELO drop due to significant SMP search overhead at bullet time controls.
However for long games & analysis it is recommended to turn HT on and manually set the number of processes (UCI option "Threads")
to equal the number of logical CPUs.

On the automated tuning (technical):
Tuning is done in two steps:
1. an estimate of the gradient is calculated;
2. line search is performed.
SPSA (just move a bit along the gradient direction instead of performing a line search) may also be quite effective,
though it's difficult to pick good parameters.
Fairly simple exit criteria are used for the line search. I don't employ SPRT because it's guaranteed to be optimal only for
elo=elo1 vs elo=elo0 hypothesis, whereas in this case we are interested in elo>elo0 vs elo<=elo0 test. However it's still possible
for SPRT (or some other test) to be better for some elo-elo0 values. Shouldn't be very difficult to run a numerical comparision
between various tests.


Best wishes,
Norbert



Let me be sure I have this correct. On my 6-Core i7, I run all matches with HT disabled. If I want to run a 4-Core match and set this Gull to use 4 Cores, will it only use 4 of the 6 true Cores available- or will it attempt to grab all 6 and use them. You would be surprised at the number of engines that attempt this- so I have to run them on one of my i5 4-Core systems, because they cannot take what aint there.

I ask because at the moment I do not have time to install it and check myself. One last thing- would you just make a ballpark guess as to the elo increase for this version. Trust me- I hold you to nothing. I WILL NOT be coming back and saying, "But you told me blah, blah."

Talking to others here- not you. Please don't be telling me how I need to use HT- not interested.



All the best to you Norbert,

george

SzG
Posts: 2447
Joined: Fri Mar 10, 2006 6:20 am
Location: Szentendre, Hungary

Re: GullChess 2.1

Post by SzG » Tue Jun 18, 2013 8:11 am

velmarin wrote:
Norbert Raimund Leisner wrote:but Gull 2.1 exe does not run on Windows XP 3 (32-bit)

error message (translated from German) :
Gull 2.1 exe is not an allowed win32-application

Probe now, ThinkingALot in Inmortal Forum:

Code: Select all

Re&#58; Gull
I recompiled the 32-bit version so it doesn't require any additional DLLs now. Gull 2.1.zip file on sourceforge is updated.
Last edited by ThinkingALot; Today at 07&#58;54.
It does not run at all. Dependency Walker says:

Error: At least one module has an unresolved import due to a missing export function in an implicitly dependent module.
Warning: At least one module has an unresolved import due to a missing export function in a delay-load dependent module.
Gabor Szots

CCRL testing group

Norbert Raimund Leisner
Posts: 1445
Joined: Tue May 20, 2008 2:57 pm
Location: Augsburg - Germany

Re: GullChess 2.1

Post by Norbert Raimund Leisner » Tue Jun 18, 2013 8:47 am


Norbert Raimund Leisner
Posts: 1445
Joined: Tue May 20, 2008 2:57 pm
Location: Augsburg - Germany

Re: GullChess 2.1

Post by Norbert Raimund Leisner » Tue Jun 18, 2013 8:49 am


SzG
Posts: 2447
Joined: Fri Mar 10, 2006 6:20 am
Location: Szentendre, Hungary

Re: GullChess 2.1

Post by SzG » Tue Jun 18, 2013 9:55 am

Norbert Raimund Leisner wrote:InterlockedCompareExchange64 http://msdn.microsoft.com/en-us/library ... s.85).aspx was not found in kernel 32.dll

http://sourceforge.net/projects/gullchess/
Very good. By its name it seems to be a 64-bit something. So the 32-bit compile is not fully 32-bit.
Gabor Szots

CCRL testing group

User avatar
Sylwy
Posts: 3358
Joined: Fri Apr 21, 2006 2:19 pm
Location: IASI (Romania) - the historical capital of MOLDOVA

Re: GullChess 2.1

Post by Sylwy » Tue Jun 18, 2013 10:06 am

SzG wrote:
So the 32-bit compile is not fully 32-bit.
Now it's OK !

http://sourceforge.net/projects/gullchess/

SilvianR :wink:

SzG
Posts: 2447
Joined: Fri Mar 10, 2006 6:20 am
Location: Szentendre, Hungary

Re: GullChess 2.1

Post by SzG » Tue Jun 18, 2013 10:14 am

Sylwy wrote:
SzG wrote:
So the 32-bit compile is not fully 32-bit.
Now it's OK !

http://sourceforge.net/projects/gullchess/

SilvianR :wink:
Yes, finally the third download works.
Gabor Szots

CCRL testing group

Post Reply