New Cerebellum Library

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Dann Corbit
Posts: 9993
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: New Cerebellum Library

Post by Dann Corbit » Wed Mar 27, 2019 7:11 pm

Here is a link for the PGN built using the technique described using polyglot:

https://drive.google.com/open?id=19IhT8 ... 6PoJw5DoRC

Even though it is enormously larger, it looks to me like both colors were truncated at around 75,000 games.
So I guess the conversion is not complete.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

clumma
Posts: 177
Joined: Fri Oct 10, 2014 8:05 pm
Location: Berkeley, CA

Re: New Cerebellum Library

Post by clumma » Thu Mar 28, 2019 2:49 am

Dann Corbit wrote:
Wed Mar 27, 2019 7:11 pm
Here is a link for the PGN built using the technique described using polyglot:

https://drive.google.com/open?id=19IhT8 ... 6PoJw5DoRC

Even though it is enormously larger, it looks to me like both colors were truncated at around 75,000 games.
So I guess the conversion is not complete.
Thanks again.

There's a lot of redundancy in such an approach, and it seems we should avoid the step of generating pseudo-games if possible.

We want something that can parse the polyglot format directly and display a tree. It would need to have a move generator of some kind.

For the demo on Zipproth's website, the GUI looks up FENs in a JSON file. I've placed the last version of that file I was able to archive here:

https://we.tl/t-J10Sv0j8kN

(File is 2.8 MB. The host is Wetransfer and is not too annoying.)

The polyglot format spec is here:

http://hgm.nubati.net/book_format.html

Brainfish needs the weights, right? Looking at Cerebellum_Light_Poly.bin in a hex editor just now, all the weights appear to be set to 00 FF!

I know Brainfish has an option to randomly play 2nd-best moves... it needs access to the weights to do that. Am I parsing this file wrong?

-Carl

Ferdy
Posts: 4077
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: New Cerebellum Library

Post by Ferdy » Thu Mar 28, 2019 4:25 am

clumma wrote:
Thu Mar 28, 2019 2:49 am
Dann Corbit wrote:
Wed Mar 27, 2019 7:11 pm
Here is a link for the PGN built using the technique described using polyglot:

https://drive.google.com/open?id=19IhT8 ... 6PoJw5DoRC

Even though it is enormously larger, it looks to me like both colors were truncated at around 75,000 games.
So I guess the conversion is not complete.
Thanks again.

There's a lot of redundancy in such an approach, and it seems we should avoid the step of generating pseudo-games if possible.

We want something that can parse the polyglot format directly and display a tree. It would need to have a move generator of some kind.

For the demo on Zipproth's website, the GUI looks up FENs in a JSON file. I've placed the last version of that file I was able to archive here:

https://we.tl/t-J10Sv0j8kN

(File is 2.8 MB. The host is Wetransfer and is not too annoying.)
That one would be easier to parse and with lots of info already, but of course this is for demo purposes file is not too big. Only the author can produce that json data.
The polyglot format spec is here:

http://hgm.nubati.net/book_format.html

Brainfish needs the weights, right? Looking at Cerebellum_Light_Poly.bin in a hex editor just now, all the weights appear to be set to 00 FF!

I know Brainfish has an option to randomly play 2nd-best moves... it needs access to the weights to do that. Am I parsing this file wrong?
In Cerebellum_Light_Poly.bin there are positions where it offers 2 moves. Example from startpos.

Image

And after e4 we have.

Image

There's a lot of redundancy in such an approach, and it seems we should avoid the step of generating pseudo-games if possible.
Dumping of book moves indeed has lots of book lines. This is good as all transpositions will be captured. What probably is needed is to make a selective extraction, for example we want a nimzo-indian, we can use pgn-extract and tell it to output only those positions that starts with d4 Nf6 c4 e6 Nc3 Bb4 using the dump pgn as the source.

Ferdy
Posts: 4077
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: New Cerebellum Library

Post by Ferdy » Thu Mar 28, 2019 1:18 pm

pohl4711 wrote:
Tue Mar 26, 2019 7:37 am
Last night, Thomas Zipproth released a new Cerebellum Library for Brainfish. First new release since the middle of 2018(!)

Download Cerebellum and latest Brainfish here:

http://www.zipproth.de/#Brainfish_download


I started the testrun... Result not before next Monday on my website (https://www.sp-cc.de).
Stay tuned!
What is the average analysis time per position used in cerebellum?

JollyJoker
Posts: 3
Joined: Sat Feb 16, 2019 7:43 pm
Full name: Lennart Qvarnström

Re: New Cerebellum Library

Post by JollyJoker » Thu Mar 28, 2019 3:31 pm

Are these evaluated using Stockfish? It would be interesting to see a comparison with a Leela version. Even more interesting if the AplhaZero team made a book with a trillion positions or so :D

Ferdy
Posts: 4077
Joined: Sun Aug 10, 2008 1:15 pm
Location: Philippines

Re: New Cerebellum Library

Post by Ferdy » Thu Mar 28, 2019 3:48 pm

JollyJoker wrote:
Thu Mar 28, 2019 3:31 pm
Are these evaluated using Stockfish?
I think by Brainfish with Stockfish latest dev.

Dann Corbit
Posts: 9993
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: New Cerebellum Library

Post by Dann Corbit » Thu Mar 28, 2019 7:56 pm

Of the 669,691 distinct board positions in the extracted part of the subset of the Cerebellum library, I have all positions analyzed, but the vast majority are weakly analyzed (my definition is less than 36 plies).

654,569 are less than 36 plies
616,896 are less than 30 plies
It will take more than a year to analyze all of them to 36 plies or better.

On the other hand, I may have side table analysis for a majority of them. (All the contests like CCRL, CEGT, TCEC, etc. are decomposed into analyzed EPD records in tables named after the contests.). So the workload might not really be that daunting.

Some of the contests (e.g. Sedat's book contests) do not have analysis embedded, but only the PGN {and in a few cases time used}. So those are good for statistics, but not helpful for engine analysis.

To me, "make this move 100%" or "make this move 33%" is not all that helpful. I want to know why.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

Dann Corbit
Posts: 9993
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: New Cerebellum Library

Post by Dann Corbit » Thu Mar 28, 2019 8:02 pm

Here is a link to all of the distinct positions in the extracted part of the subset of the Cerebellum library:
https://drive.google.com/open?id=1HgwCF ... Xga02YfNV0

Remember, that this is "Cerebellum Light" so not all of the stuff is inside it.

As to the missing evaluation data, keep in mind that they switched to polyglot book format, which does not have evaluation data in it.
I guess that the polyglot book is the final output, but the intermediate book format will have the evaluation data and the real statistics in it.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.

Post Reply