TalkChess.com

Is it possible for nowadays strong engines that an engine can write its own opening book? We tried the almighty Stockfish 10. From 82,000 human games......

....

http://rebel13.nl/home/whatever.html

With the help of the new MEA

Thomas Zipproth did it already. With Stockfish and does it with lc0.

https://zipproth.com/Brainfish/download

Does the Cerebellum Library simply consists of a lot of calculated positions? No, the Cerebellum Library does a recalculation of every score in the Library. That recalculation is done with a graph Algorithm, so that every Score in a node is consistant with the score of it’s best successor. Furthermore the Cerebellum Library is able to handle position repetitions by with re-visiting some nodes and then playing different moves.

Testing results on my website:
https://www.sp-cc.de

Regards - Stefan (SPCC)

pohl4711 wrote: ↑Tue Jun 18, 2019 3:14 pm Thomas Zipproth did it already. With Stockfish and does it with lc0.

https://zipproth.com/Brainfish/download

Does the Cerebellum Library simply consists of a lot of calculated positions? No, the Cerebellum Library does a recalculation of every score in the Library. That recalculation is done with a graph Algorithm, so that every Score in a node is consistant with the score of it’s best successor. Furthermore the Cerebellum Library is able to handle position repetitions by with re-visiting some nodes and then playing different moves.

How fast is its recalculation given half a million positions?

Regarding repetition, when there is repetition, the alternatives are usually inferior which may lead to a losing position eventually. Do you mean that cerebellum will not save moves that may lead to repetitions?

I think it's worth noting that I had tested CerebellumLight bin book against all the bin books that can be downloaded from Rebel's site (including some humongous one by Dann Corbit and all related versions, and the best effort by Jeroen Noomen), and CerebellumLight destroyed them all with ease.

It was an embarrasing "no contest".

However, it's easy nowadays to download a public book that easily defeats CerebellumLight (because all its weaknesses have been found and exploited.)

People have been dedicating lately to try to create the ultimate opening book, and for this they download all the public books (usually on CTG and bin formats, but there's really strong bix books as well), find their holes and release something stronger (this is why commercial books have no value; someone has already bought them and busted all their lines.)

It seems the life of a public book is about 2 weeks, that's the time the latest thing includes the lines to abuse all their lines and destroy it.

But these are the most exciting times for opening theory, today an InfinityChess player was showing how strong the Napoleon opening (1.e4 e5 2.Qf3) can be, against best Stockfish derivatives and Leelas running in strong GPUs.

Which shows opening theory is still in diapers.

Ovyron wrote: ↑Tue Jun 18, 2019 11:30 pm I think it's worth noting that I had tested CerebellumLight bin book against all the bin books that can be downloaded from Rebel's site (including some humongous one by Dann Corbit and all related versions, and the best effort by Jeroen Noomen), and CerebellumLight destroyed them all with ease.

It was an embarrasing "no contest".

However, it's easy nowadays to download a public book that easily defeats CerebellumLight (because all its weaknesses have been found and exploited.)

Which is the point, meaning pointless. Any book can be hacked once it is publicly available. Attached a Polyglot opening book I made about a year ago of just 22Kb (1.345 positions) that beats the a CL book of 2018. There is no fun in that kind of competition.

pohl4711 wrote: ↑Tue Jun 18, 2019 3:14 pm Thomas Zipproth did it already. With Stockfish and does it with lc0.

I am aware but I think you do realize this is a new approach.

The CL books are great but it is limited to comp-comp and while I can imagine people have fun competing it does not contribute much to human chess because the extremely low varying of the book. The last time I ran a 5000 bullet match between 2 equal CL books I got about 4880 doubles and only 120-125 different openings, 125/5000. This is of course part of its strength, I get that, but I wonder how useful it is for the average chess player.

Compare this with the small (but varied) freeware books, when I play PD vs PD I get 2800/5000 unique openings, not 120-125.

Nice project Ed (and Ferdy).

Rebel wrote: ↑Wed Jun 19, 2019 9:55 am The last time I ran a 5000 bullet match between 2 equal CL books I got about 4880 doubles and only 120-125 different openings, 125/5000. This is of course part of its strength, I get that, but I wonder how useful it is for the average chess player.

What do you mean by "doubles"? Do you mean the entire game was the same?

jp wrote: ↑Wed Jun 19, 2019 10:05 am

Rebel wrote: ↑Wed Jun 19, 2019 9:55 am The last time I ran a 5000 bullet match between 2 equal CL books I got about 4880 doubles and only 120-125 different openings, 125/5000. This is of course part of its strength, I get that, but I wonder how useful it is for the average chess player.

What do you mean by "doubles"? Do you mean the entire game was the same?

No, just the book lines.

Rebel wrote: ↑Wed Jun 19, 2019 1:19 am There is no fun in that kind of competition.

People find enjoyment in being the best in something that is hard, and believe me, making a book that is the best today is really hard, the competition is fierce and people are coming up with books that destroy everything in the past and get really though when anything else (not from 2 weeks from the future) tries to beat them.

Opening theory is making the fastest progress I've ever seen (like, did you know white's Giuoco Piano has been busted? By black? That's amazing! I can beat in my 10 years old computer someone with the latest thing if they play the Italian as white) and people are missing it because they want "variety" (even though what they call variety, I call "irrelevant, weak, suboptimal chess lines".)