corres wrote:Values displayed by engines reflect the opinion of developers only. Their opinion is based on practical games without theoretical evidence.
This strikes me as a very strange claim. Engine evals are associated with winning probabilities (through logistic curve). We see that in game statistics and Monte Carlo analysis. They are opinions of supernaturally strong chess players, not engine developers. At least not since Rajlich. The era of developer opinions, during which all progress came from hardware improvement, is safely behind us.
[quote="clumma"]
This strikes me as a very strange claim. Engine evals are associated with winning probabilities (through logistic curve). We see that in game statistics and Monte Carlo analysis. They are opinions of supernaturally strong chess players, not engine developers. At least not since Rajlich. The era of developer opinions, during which all progress came from hardware improvement, is safely behind us.
-Carl[/quote]
But you want to use the results of Stockfish and not the sentences of Magnus Carlsen, for e.g.
Naturally, I believe that you can make a super strong opening book.
The success depends on money, work of experts and time of machines.
But to solve the problem of chess game you ought to produce the 32 men database. With the current technology it is an impossible task.
drj4759 wrote:Cerebellum is arguably the strongest opening book ever. Its strength lies in the memorized best value score of a given move. If it has not seen that move before then it behaves like a normal Stockfish chess engine/book.
With absolute certainty, the unchanging first move of Cerebellum as white is e4 and its favorite variation is the Giuoco Piano. With black against e4, it plays the constant e5.
So, it is possible to find ways to neutralize by thorough preparation and accumulation of the best move sequence against these. This is both a strength by having a large collection of store best move and at the same time a weakness as other book makers could assimilate that large best move collection and then use it against Cerebellum with the injection of more best moves that has not been memorized.
In theory, chess games should end a draw with correct play. Good opening books makes it easier to play correctly. The best opening book actively maintained thus far is Cerebellum but may not sustain its perceived strength when somebody else will have the focus to create an alternative book that will neutralize it.
Unfortunately, the best reply on 1.e4 is not 1...e5, but 1...c5.
Also, the giuoco (goddam wops, how can they insert 3 vowels one after another ) piano is much inferior for white than the Ruy Lopez and a range of other sequences.
Biggest downside of this project though is that this book, no matter the hours spent analysing and the depth reached, will be easily beaten by any book created by some 200 elo stronger than current SF engine in ..... 5 minutes.
What kind of nonsense do you write is really beyond me. First Cerebellum gives the best response to e4 as French, i.e. e6, not e5, then e5 and than Sicilian, the thing is the differences are so small that all 3 are practically equal.
Seams you really have no clue how Cerebellum works. The analysis they did is really deep, and I am sure it took quite an effort and burned quite some electricity. Basically you propagate all the way to the last move of the line and then do full depth search at leaves basically, then you back-propagate the evaluation score to the root. That way you essentially get 55-60 depth at root. That is an immense depth, something you can't reach in correspondence even with the strongest machine at the world. Considering that every additional ply is at least 30-40Elo that is at least 800Elo more than typical 1minute search at root on strong quad core.
No pathetic human chess opening theory developed for last couple of centuries can be a match for such a deep search by the strongest engine on the planet.
Thinking that you know better than Stockfish at depth 60 at root is just plain delusional.
drj4759 wrote:The opening book that can neutralize Cerebellum is here. I invite you to visit the ChessOwl site later where I post the result of the ongoing tournament between Brain Fish/Cerebellum vs. Fauzi 2.2 and Fishbook Dynamic 161016. Partial results showed the overall superiority of Cerebellum but lags behind Fishbook.
Again someone who doesn't understand what Cerebellum is. Cerebellum is not Brainfish. Brainfish is Stockfish with a very small and crippled subset of Cerebellum as opening book. It is only 1 to 2 moves per ply without score value or depth value. That is basically useless. Size of Cerebellum book is at least 4-5 orders of magnitude bigger than Brainfish book and difference in strength is probably even couple of hundreds of Elo.
drj4759 wrote:The opening book that can neutralize Cerebellum is here. I invite you to visit the ChessOwl site later where I post the result of the ongoing tournament between Brain Fish/Cerebellum vs. Fauzi 2.2 and Fishbook Dynamic 161016. Partial results showed the overall superiority of Cerebellum but lags behind Fishbook.
Again someone who doesn't understand what Cerebellum is. Cerebellum is not Brainfish. Brainfish is Stockfish with a very small and crippled subset of Cerebellum as opening book. It is only 1 to 2 moves per ply without score value or depth value. That is basically useless. Size of Cerebellum book is at least 4-5 orders of magnitude bigger than Brainfish book and difference in strength is probably even couple of hundreds of Elo.
That sucks. When do we get the real Cerebellum book then?
drj4759 wrote:The opening book that can neutralize Cerebellum is here. I invite you to visit the ChessOwl site later where I post the result of the ongoing tournament between Brain Fish/Cerebellum vs. Fauzi 2.2 and Fishbook Dynamic 161016. Partial results showed the overall superiority of Cerebellum but lags behind Fishbook.
Again someone who doesn't understand what Cerebellum is. Cerebellum is not Brainfish. Brainfish is Stockfish with a very small and crippled subset of Cerebellum as opening book. It is only 1 to 2 moves per ply without score value or depth value. That is basically useless. Size of Cerebellum book is at least 4-5 orders of magnitude bigger than Brainfish book and difference in strength is probably even couple of hundreds of Elo.
That sucks. When do we get the real Cerebellum book then?
It's all in FAQ:
Is the tool for creating, expanding, viewing and managing the Library available?
It will be available in form of the standalone chess GUI Sirius, including additional features. It will probably published around November 2016 and will cost around 20,- €. Brainfish itself will always stay free.
APassionForCriminalJustic wrote:That sucks. When do we get the real Cerebellum book then?
It's all in FAQ:
Is the tool for creating, expanding, viewing and managing the Library available?
It will be available in form of the standalone chess GUI Sirius, including additional features. It will probably published around November 2016 and will cost around 20,- €. Brainfish itself will always stay free.
The problem I see is that it will probably be in the format of Lomonosov, so you won't be able to actually download the full library but instead through their GUI probe the positions and probably couple it with your favourite engine and also add your own positions stored locally. Ofc downloading the full library would beat the purpose of selling the GUI, i.e. they wouldn't earn anything. If you think about it, it is logical that the project should at least compensate for the burned electricity/computer power they used, so having Brainfish as trial library coupled with SF is a nice introduction into commercial product.
Milos wrote:The problem I see is that it will probably be in the format of Lomonosov, so you won't be able to actually download the full library but instead through their GUI probe the positions and probably couple it with your favourite engine and also add your own positions stored locally. Ofc downloading the full library would beat the purpose of selling the GUI, i.e. they wouldn't earn anything.
Anyone who knows how to use AWS can replicate their work in a day for about $500. So I suggest they don't charge more than that.
Milos wrote:The problem I see is that it will probably be in the format of Lomonosov, so you won't be able to actually download the full library but instead through their GUI probe the positions and probably couple it with your favourite engine and also add your own positions stored locally. Ofc downloading the full library would beat the purpose of selling the GUI, i.e. they wouldn't earn anything.
Anyone who knows how to use AWS can replicate their work in a day for about $500. So I suggest they don't charge more than that.
-Carl
You are really greatly underestimating the effort.
For 500$ you can take around 10k core hours on the fastest core, i.e. 10k*3600s*2Mnps = 72Tnodes. Depths they achieve on the leaf positions on Cerebellum require at least 1Gnodes and these are the shallowest. Since branching factor is high leaf positions are like 90% of total number of positions. And there are 4.5M positions, i.e. around 4M leaf positions. Lets assume leaf positions are only searched to 200Mnodes (i.e. only 100s of search time per position) which is quite an underestimation considering depths achieved at leaf positions. 4.5M*200Mnodes~1000Tnodes. So you just made a 14 fold underestimation while actual cost on AWS would be something like 7k$ or more realistically around 10k$.
I can't find the post where Zipproth gave the $2000 of electricity figure. It's possible that number was for Cerebellum light.
I did find this post where he states Cerebellum (I think the full version) represents 2 years of time on 2 quad-core machines. To do it in a day we need ~ 1500 c4.xlarge instances at $0.21/hr = $7500.
) piano is much inferior for white than the Ruy Lopez and a range of other sequences.