FEN database (287 millions posiitions)

mathmoi · Post by **mathmoi** » Sat May 04, 2013 3:39 pm

Hi,

Back in 2005 I posted on rec.games.chess.computer that I had a FEN position database containing millions of positions. Since then a couple of persons emailed me asking to get it usually to use in some research or university projects.

I just updated the database and put it on my website if anyone is interested :

http://mathieupage.com/?p=65

Adam Hair · Post by **Adam Hair** » Sat May 04, 2013 8:28 pm

Thank you very much, Mathieu! I have a few hundred thousand positions in epd format that I have been using. 286 million is much better.

mathmoi · Post by **mathmoi** » Sat May 04, 2013 9:59 pm

You're welcome.

By curiousity, what do you use thoses for?

Adam Hair · Post by **Adam Hair** » Sat May 04, 2013 11:28 pm

I used them to compare Gaviota's evaluation to the evaluations from a group of stronger engines. I collected the depth 4 evaluations from all of the stronger engines, determined the mean and standard deviation of the evals for each position, then compared Gaviota's eval to the group mean. For any position that was several deviations different from the mean, I studied Gaviota's behavior for that position in more detail and sent my findings to Gaviota's author, Miguel Ballicora. Miguel has been able to correct some undesired behavior in Gaviota from this information.

mathmoi · Post by **mathmoi** » Sat May 04, 2013 11:41 pm

That's an interesting way to find weak spot in an engine's evaluation.

Adam Hair · Post by **Adam Hair** » Sat May 04, 2013 11:50 pm

The idea was Miguel's. It looked like a good idea to me, so I went forward with it.

Adam Hair · Post by **Adam Hair** » Sat May 04, 2013 11:58 pm

By the way, we have also found a few flaws with the search with this method.

Desperado · Post by **Desperado** » Sun May 05, 2013 12:04 pm

Thx Mathieu !

I spent today some of my rare time today to implement some functions to get use of this data.
Then suddenly out of the blue sky Nemo was crashing and i need at least 30 min to find out what was going on.

Warning for UCI followers:

Chess Interfaces and GUIs will never send Mate and Stalemate positions
to the engine. So, not expecting to get a Mate Position from fen, Nemo never handled such case.
If not already done, one should catch these cases at the root, which is normally not necessary for uci implementation. Somthing like will do the job.

Code: Select all

    // Ignore Mate / Stalemate &#40;especially injected by Fen Test DB&#41;
    if&#40; mli->size == 0 )
    &#123; return INCHECK ? mateIn&#40;0&#41; &#58; STALEMATE; &#125;

Hopefully this saves some unecessary debugging time.

Michael

mathmoi · Post by **mathmoi** » Sun May 05, 2013 4:27 pm

Hi Michael,

You're welcome. I'm happy it can be useful.

mp

Tom Likens · Post by **Tom Likens** » Sun May 12, 2013 5:02 pm

Adam Hair wrote:I used them to compare Gaviota's evaluation to the evaluations from a group of stronger engines. I collected the depth 4 evaluations from all of the stronger engines, determined the mean and standard deviation of the evals for each position, then compared Gaviota's eval to the group mean. For any position that was several deviations different from the mean, I studied Gaviota's behavior for that position in more detail and sent my findings to Gaviota's author, Miguel Ballicora. Miguel has been able to correct some undesired behavior in Gaviota from this information.

Adam,

This is a really interesting idea. How did you automate this?

regards,
--tom

FEN database (287 millions posiitions)

FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)

Re: FEN database (287 millions posiitions)