Brainfish, a new concept of a chess engine

[corres]

As I see Cerebellum Library is an opening book loaded into RAM. So the reaching time is very low. This is important for the very fast test matches.
But there are some question about it:
-How much RAM does it need?
-How many engines can be connected to the interface of the library?
-Where get from the value of an opening position?
The last is the main question because so many engines so many values.

[Lyudmil Tsvetkov]

So, the commercial project is not related to SF or any other engine, but you can do the calculations with SF or any other engine?? Are not those 2 statements the same? You are not using an engine while you are using an engine?

So, the Cerebellum library is a SF-calculated library after all?

So, please kindly explain where is the difference between SF and Brainfish, as well as between SF-calculated library and Cerebellum library.

Maybe you would like to release only a GUI after all? It might have its merits, I have not investigated that matter.

Sorry, maybe my english was not the best at this point or I was not very clear, I explain again:

The Commercial Product will be:

1. The Gui with Cerebellum Library handling
2.) The Cerebellum Library itself, which does not mean an already calculated book, but a tool (Cerebellum) to make such a book using an engine of your choice.
3.) An adapter to connect a Cerebellum calculated book with an engine for playing.
4.) The book created with the Cerebellum Library using Stockfish as engine as an additional benefit.

So perhaps the misunderstanding was that the Cerebellum Library is not an already calculated book, it's the tool to create and use such a book. It's in fact a software library to do that.

I hope now it is more clear.

Thomas

Excellent!

From the points you enumerated above:

1. No matter how appealing a GUI migth be, it is very hard to call it innovative.
2. A tool for creating a book using an engine of one's own choice is a nice thing, but hardly innovative. As far as I know, most engines are able to create books automatically.
3. Programmers migth advise if an adapter connecting a book to an engine is a very innovative concept, to me it does not quite seem this way.
4. We already discussed the book itself.

So, the point is, where is the innovation up there?
Because, it is publicised like that, something that has never been done.

Please, do not get me wrong, I appreciate your effort, but, let's call things with their names. If you state that you have created some tools for building a custom, engine-based book in an easy manner within its own GUI, I fully endorse that and, presumably, many other people will do the same. But why claim that this is something more than a book-building tool?

It is indeed a book-building tool, whose product is later connected to the engine.
Concerning the quality of the book built, I hope it has its merits, additional time spent on calculating lines always helps, but it would be interesting to match SF with your library with SF with a book of the same size/depth, based on GM games, for example. If Brainfish is using library with 50 plies depth, then standard SF shall also use a book with 50-plies depth.

And you know what, I bet that there will be no elo increase over standard SF. For the simple reason, that GM lines still vastly outperform SF-built lines. Please try that, and then after reporting the result, decide how to publicise your project.

Matching SF with your library with 50-plies depth vs SF with 5-10 plies book certainly leaves standard SF at an enormous disadvantage, especially at 1' TC, I bet after executing all its precalculated moves, SF-Cerebellum already has very significant time advantage.

Again, I appreciate your effort, but let's call things with their names.

[Ozymandias]

Concerning the quality of the book built, I hope it has its merits, additional time spent on calculating lines always helps, but it would be interesting to match SF with your library with SF with a book of the same size/depth, based on GM games, for example. If Brainfish is using library with 50 plies depth, then standard SF shall also use a book with 50-plies depth.

And you know what, I bet that there will be no elo increase over standard SF. For the simple reason, that GM lines still vastly outperform SF-built lines. Please try that, and then after reporting the result, decide how to publicise your project.

Already done, googling a result shared by Stefan Pohl, I found in the subsequent page, this one:

Code: Select all

Now ASM receives TopGM_8move.pgn....what a difference just an 8 move book makes !!
-16 Elo for Brainfish

# PLAYER              :  RATING  ERROR  PLAYED   (%)  D(%)
   1 SF 160722ASMPD      :    3208      7    1000  52.3  61.0
   2 Brainfish 160724    :    3192      7    1000  47.7  61.0

[Dann Corbit]

Concerning the quality of the book built, I hope it has its merits, additional time spent on calculating lines always helps, but it would be interesting to match SF with your library with SF with a book of the same size/depth, based on GM games, for example. If Brainfish is using library with 50 plies depth, then standard SF shall also use a book with 50-plies depth.

And you know what, I bet that there will be no elo increase over standard SF. For the simple reason, that GM lines still vastly outperform SF-built lines. Please try that, and then after reporting the result, decide how to publicise your project.

Already done, googling a result shared by Stefan Pohl, I found in the subsequent page, this one:

Code: Select all

Now ASM receives TopGM_8move.pgn....what a difference just an 8 move book makes !!
-16 Elo for Brainfish

# PLAYER              :  RATING  ERROR  PLAYED   (%)  D(%)
   1 SF 160722ASMPD      :    3208      7    1000  52.3  61.0
   2 Brainfish 160724    :    3192      7    1000  47.7  61.0

And yet you can take a billion rows of computer generated data and make a very deep and reliable book automatically. You can also extend it into special openings and calculate deep traps if you like.

I remember some time ago {back when a computer would think all night for 18 plies and cough up a crappy answer in the Evans Gambit}, Vincent told me that computer analysis would never be able to compete with human generated books. Since the performance of these two books are practically identical and one is completely computer generated, I would say that day has already arrived. Consider also that this is a truncated version of the book.

What the tool really provides (so far as I can tell) is an easy way to minimax a big volume of computer data, and to deliver that data in a standards based way to a computer engine.

Too early to say how big the benefit will be, but it is also too early to poo-poo the idea as not worthy of examination.

[Ozymandias]

I remember some time ago {back when a computer would think all night for 18 plies and cough up a crappy answer in the Evans Gambit}, Vincent told me that computer analysis would never be able to compete with human generated books. Since the performance of these two books are practically identical and one is completely computer generated, I would say that day has already arrived. Consider also that this is a truncated version of the book.

What the tool really provides (so far as I can tell) is an easy way to minimax a big volume of computer data, and to deliver that data in a standards based way to a computer engine.

Too early to say how big the benefit will be, but it is also too early to poo-poo the idea as not worthy of examination.

You would need to ask the guy testing, but the book might be the one used is Fishtest, which was extensively engine-checked.

I think the idea (the software) is valid, but the implementation (he current book) may be far from optimal.

[Dann Corbit]

I remember some time ago {back when a computer would think all night for 18 plies and cough up a crappy answer in the Evans Gambit}, Vincent told me that computer analysis would never be able to compete with human generated books. Since the performance of these two books are practically identical and one is completely computer generated, I would say that day has already arrived. Consider also that this is a truncated version of the book.

What the tool really provides (so far as I can tell) is an easy way to minimax a big volume of computer data, and to deliver that data in a standards based way to a computer engine.

Too early to say how big the benefit will be, but it is also too early to poo-poo the idea as not worthy of examination.

You would need to ask the guy testing, but the book might be the one used is Fishtest, which was extensively engine-checked.

I think the idea (the software) is valid, but the implementation (he current book) may be far from optimal.

If we can feed the book generator our own data, I know where to get a big pile of it.

[Thomas Zipproth]

Already done, googling a result shared by Stefan Pohl, I found in the subsequent page, this one:

Code: Select all

Now ASM receives TopGM_8move.pgn....what a difference just an 8 move book makes !!
-16 Elo for Brainfish

# PLAYER              :  RATING  ERROR  PLAYED   (%)  D(%)
   1 SF 160722ASMPD      :    3208      7    1000  52.3  61.0
   2 Brainfish 160724    :    3192      7    1000  47.7  61.0

Yes, I had those results too with an 8 move gm book. The reason is that the variability after playing 8 moves from a gm book is so big, that most of the those positions are not in the book, even with 4.4 Million positions in it. So that is not unexpected. Most positions in the book are from engine tournaments, and some from the most played positions by gm's.

[Lyudmil Tsvetkov]

Concerning the quality of the book built, I hope it has its merits, additional time spent on calculating lines always helps, but it would be interesting to match SF with your library with SF with a book of the same size/depth, based on GM games, for example. If Brainfish is using library with 50 plies depth, then standard SF shall also use a book with 50-plies depth.

And you know what, I bet that there will be no elo increase over standard SF. For the simple reason, that GM lines still vastly outperform SF-built lines. Please try that, and then after reporting the result, decide how to publicise your project.

Already done, googling a result shared by Stefan Pohl, I found in the subsequent page, this one:

Code: Select all

Now ASM receives TopGM_8move.pgn....what a difference just an 8 move book makes !!
-16 Elo for Brainfish

# PLAYER              :  RATING  ERROR  PLAYED   (%)  D(%)
   1 SF 160722ASMPD      :    3208      7    1000  52.3  61.0
   2 Brainfish 160724    :    3192      7    1000  47.7  61.0

And yet you can take a billion rows of computer generated data and make a very deep and reliable book automatically. You can also extend it into special openings and calculate deep traps if you like.

I remember some time ago {back when a computer would think all night for 18 plies and cough up a crappy answer in the Evans Gambit}, Vincent told me that computer analysis would never be able to compete with human generated books. Since the performance of these two books are practically identical and one is completely computer generated, I would say that day has already arrived. Consider also that this is a truncated version of the book.

What the tool really provides (so far as I can tell) is an easy way to minimax a big volume of computer data, and to deliver that data in a standards based way to a computer engine.

Too early to say how big the benefit will be, but it is also too early to poo-poo the idea as not worthy of examination.

What is poo-poo?

Please note that in the reported test, asm had 8-move/16-plies-deep book pitted vs 50-plies book.
Arrange a new competition with a deeper book along the size of CRB, or, alternatively, limit the depth of CRB to 16 plies, and then testers would be even more surprised, because already asm will lead by 70 points, the same amount before the tables were turned on the book.

At the present point in time, not sure how many seconds since January 1st, 1970, top engines are still far far apart from optimal theoretical lines.

[Ozymandias]

If we can feed the book generator our own data, I know where to get a big pile of it.

Does his name start with an "r"?

[Thomas Zipproth]

So, the point is, where is the innovation up there?
Because, it is publicised like that, something that has never been done.

And you know what, I bet that there will be no elo increase over standard SF. For the simple reason, that GM lines still vastly outperform SF-built lines. Please try that, and then after reporting the result, decide how to publicise your project.

Again, I appreciate your effort, but let's call things with their names.

I see your point, at least in parts, but let me explain why it was advertised as "Brainfish, a new concept of a chess engine", and what the innovation is.

What is the function of a Brain? Besides some other things, it creates new knowledge out of existing knowledge. That is exactly what the library does. It takes all the evaluation scores created by an engine for every position in the library, and calculates new scores out of it with it's graph algorithm. Together with that the main lines are created, and new knowledge about transistions and possible forced repetitions is found and can be displayed and used for playing.

That is the innovation, at least I don't know another product or free library that does it that way.

The second point which I quoted from you I do not understand. You say there will be no elo increase over Standard SF because GM lines still vastly outperform SF-built lines. Do you mean compared to SF with a book of GM lines?