Random Books v4.0 for Shredder, Fritz GUI, engines ready!

[Frank Quisinsky]

Hi there,

Dokumentation in German
Translate it with Google, link in front of the page.

http://www.amateurschach.de/master-pgn/ ... tation.htm

Big download file:
All steps to the results, Stockfish analyzes, epd files and so on
~ 7Mb, ZIP file

Little download file:
Only the two important PGN files and the same random book for Shredder (*.bkt) and Fritz GUI (*.ctg).
~ 1Mb, ZIP file

More or less ...
3.920 start positions will be played with an opening book by random. Optimal for Eng-Eng matches.

Included is a blunder database!
All blunders are collected and available in EPD and PGN. So its easy to check available opening books and to see where come the blunders.

More or less all what is in the database > 0,68

Example:
In ChessBase Mega 2010 is one game (GM vs. GM) Stockfish gave after black move 8. the evaluation: 7.0 !! This blunder I found in different newer opening books, quit clear the most useing the same GM games.

The complete Stockfish analyzes are included.
Analyze: SF 1.7.1 x64 4-Cores, 1.024 Hash, 1 minute per position!
All positions are deleated > 0,68

Easy to check after an eng-eng game a critical position:

Example:
Eng A - Eng B.
First move to calculate is white move 9.
Now Engine A give us 1.0 (normaly not possible).

Load an log file from Stockfish in an editor and search the position. You can see that for SF is the position OK.

Arena have the PGN-Random option, Shredder and Fritz GUI not have. For Arena is only the PGN file version 3.5 or 4.0 interesting.

3.5: with rarely variants
4.0: without rarely variants (1.320 games deleted)

For the Shredder and Fritz books are important that the rarely variants deleted ... or the GUis play 1. a3 / 1.d4 by random ... same count of such games.

From version 4.0 I created the opening books.

Best
Frank

[Frank Quisinsky]

Hi there,

it's now very interesting to compare engines after the book moves.

Example:

Engine A: Position 2400, evaluation 0,25
Engine B: Position 2400, evaluation -0,5

What is wrong here, a good question!
I think a lot mistakes can be found here ...

So, the EPD files from the end positions v3.5 / v4.0 are included in the in "big" donwload file (again my thanks to Norm Pollock). Easy to analyse the complete EPD which different engines or to add the evaluation from this engines in Excel files.

At the moment I try to find out which opening systems Stockfish like, Naum like and so on

Naum programmer are thinking that Naum will better play in closed positions. After my first results I can't see that.

Best
Frank

[Milos]

More or less ...
3.920 start positions will be played with an opening book by random. Optimal for Eng-Eng matches.

Included is a blunder database!
All blunders are collected and available in EPD and PGN. So its easy to check available opening books and to see where come the blunders.

More or less all what is in the database > 0,68

Example:
In ChessBase Mega 2010 is one game (GM vs. GM) Stockfish gave after black move 8. the evaluation: 7.0 !! This blunder I found in different newer opening books, quit clear the most useing the same GM games.

The complete Stockfish analyzes are included.
Analyze: SF 1.7.1 x64 4-Cores, 1.024 Hash, 1 minute per position!
All positions are deleated > 0,68

Easy to check after an eng-eng game a critical position:

Example:
Eng A - Eng B.
First move to calculate is white move 9.
Now Engine A give us 1.0 (normaly not possible).

Load an log file from Stockfish in an editor and search the position. You can see that for SF is the position OK.

Arena have the PGN-Random option, Shredder and Fritz GUI not have. For Arena is only the PGN file version 3.5 or 4.0 interesting.

3.5: with rarely variants
4.0: without rarely variants (1.320 games deleted)

For the Shredder and Fritz books are important that the rarely variants deleted ... or the GUis play 1. a3 / 1.d4 by random ... same count of such games.

From version 4.0 I created the opening books.

Hi Frank,
I have two requests and a small suggestion for you.
First could you maybe extend v4.0 from 3920 to 4000 openings by adding additional 80 less rare openings from v3.5 so we have a round number useful for testing.

Second could you maybe produce a sorted (not random) version of v4.0 book, where games are sorted according to how often they appeared in the original GM database?

A suggestion is to do another analysis pass through the openings with another strong program (Rybka 4 for example) and filter those where there is a big discrepancy between SF and other engine. For example if difference in analysis is more than 1.00.

Thank you.

[Frank Quisinsky]

Hi Milos,

to add your idea is no problem. But I have to start with version 3.5 I think.

1. Analysing v3.5 (5.120 games) again with an other engine. Rybka isn't a good example (Rybka is on place 13 only) I believe. In my statistics (first moves after openings, found the most good moves), the TOP 3 are

- Spark 0.5
- Stockfish 1.7.1
- Hiarcs 13.1

If I analyze the complete v3.5 again, I would make it with Spark 0.5 I think.

An other question isn't which is the strongest engine for such an analyze!
Interesting is ... which engine will give me the most "close to reality evaluations".

I never made such experiments to "close to reality evaluations"?
Gut feeling: Komodo 1.2

Yesterday in the evening I do that with Doch 1.3.4 (800 of this positions). Highest what I saw is 0,87 after 30 seconds analyze time, 1-Core (3 positions goes over 0,8 but with the moves after all is balanced.

That's the main problem.

I made a lot of statistics before I come to the decision to use "0,65". If I used 0,25, 0,40 and so one to many games are to detelte and more important ... complete openings don't come to this filter.

I don't know Milos ...
Which Engine I have to use for a second analyze for the complete v3.5.
Spark or Komodo ... an analyzes again need 4 days (same conditions Stockfish had).

2. In around 800.000 games I had (not only ChessBase Mega 2010) are the start positions included. I found out them with the opening key system from ChessBase 10.

But now to look in each of this position in detail ... how often comes the position on the board is a lot of work!!!

Have to load each EPD in ChessBase and to search in ChessBase Online database ... 5.120x !!

Clear, it would be better to have a file with such information but what we have to do with the information?

Milos, v. 4.0 isn't important. Only for creating the books. 3.920 positions are enough to see book-random ... all seems to be perfect for me.

But most of deleting of the 1.320 games (step 3.5 to 4.0) are brute force because I must do that to get a good book random.

Important for working in the files is v. 3.5!
To add 80 position from v3.5 to v4.0 isn't really important I think.

Questions to you:

1. Which engine should analyzing the compete 5.120 positions from version 3.5 again?

Spark 0.5 (very aggressive), Komodo 1.2 (very balanced in positional playing style) or Naum ... (an engine which found most things with more analyse time). Other ideas (please don't write now a name of a clone).

2. After this step and selected again games out with evaluation higher than 0,65 I have to search how many games of this end positions are in a database! Right?

But what we should do with this information because all positions comes from the best players the World have.

If a position one times only in the database, played in Kasparow - Karpov for an example, why I have to delete it?

With rarely I mean that the positions are bad for PGN-Random-Books.
Example 1. a3

With random 1. a3 and 1. d4 is the same.

You see, the idea to create such things are good and I can understand what you mean ... but I am sure its not possible to make it perfect (perfect --- your intention I think)

I don't know what is to do now with version 3.5!

Perhaps you have interest to work in the basis file version 3.5 and create an other random-book? I have no problems with it.

What I can do is to analyse the the EPD with 5.120 again with an second engine. Hope the next one will not write ... Frank, we should have three engines for a better selection. I think one engine is enough because Stockfish analysing with a lot of hardware power and time.

Best
Frank

Thinker can do that ... on place 4 after my statistics. Very fine engine! But Thinker don't give any information about evaluations

[Frank Quisinsky]

Hi Milos,

again, if you will work on the files.

Version 4.0 isn't important.
Most important is version 3.5 !!!!!!!!!!!!!!!!!!!!

If you have interest to add own ideas for better cleanings ...
Please use version 3.5 as basis file!

Best
Frank

Nice that you have interest and to think about the idea. With more brains we get a better result.

[Frank Quisinsky]

Hi Milos,

I added your questions and my German answers in the documentation (FAQ, Point 9).

http://www.amateurschach.de/master-pgn/ ... tation.htm

Best
Frank

[Milos]

1. Analysing v3.5 (5.120 games) again with an other engine. Rybka isn't a good example (Rybka is on place 13 only) I believe. In my statistics (first moves after openings, found the most good moves), the TOP 3 are

- Spark 0.5
- Stockfish 1.7.1
- Hiarcs 13.1

If I analyze the complete v3.5 again, I would make it with Spark 0.5 I think.

An other question isn't which is the strongest engine for such an analyze!
Interesting is ... which engine will give me the most "close to reality evaluations".

I never made such experiments to "close to reality evaluations"?
Gut feeling: Komodo 1.2

Yesterday in the evening I do that with Doch 1.3.4 (800 of this positions). Highest what I saw is 0,87 after 30 seconds analyze time, 1-Core (3 positions goes over 0,8 but with the moves after all is balanced.

That's the main problem.

I made a lot of statistics before I come to the decision to use "0,65". If I used 0,25, 0,40 and so one to many games are to detelte and more important ... complete openings don't come to this filter.

I don't know Milos ...
Which Engine I have to use for a second analyze for the complete v3.5.
Spark or Komodo ... an analyzes again need 4 days (same conditions Stockfish had).

The idea of second engine would be just to avoid some blunders of SF analysis. For example if SF gives 0.5 and other engine gives more than 0.6 different evaluation for example -0.2, something is wrong. If other engine gives over 1.0 it's even worse.
So I would propose a small rule. If other engine gives analysis that is less than 1.0 different but in opposite direction it's ok. If it gives more than +0.6 different but in the same direction it's bad.
Or to write it in a more formal way. Let's ScoreSF be evaluation of SF, and Score2 be evaluation of second engine. The opening should be reconsidered if:
abs(ScoreSF - Score2)>1 and if abs(ScoreSF)-abs(Score2)>0.5

2. In around 800.000 games I had (not only ChessBase Mega 2010) are the start positions included. I found out them with the opening key system from ChessBase 10.

But now to look in each of this position in detail ... how often comes the position on the board is a lot of work!!!

If you have the original 800.000 games database, it should be quite easy to provide you the statistics of v3.5 openings database. I could do it in quick time.

Questions to you:

1. Which engine should analyzing the compete 5.120 positions from version 3.5 again?

Spark 0.5 (very aggressive), Komodo 1.2 (very balanced in positional playing style) or Naum ... (an engine which found most things with more analyse time). Other ideas (please don't write now a name of a clone).

2. After this step and selected again games out with evaluation higher than 0,65 I have to search how many games of this end positions are in a database! Right?

But what we should do with this information because all positions comes from the best players the World have.

If a position one times only in the database, played in Kasparow - Karpov for an example, why I have to delete it?

With rarely I mean that the positions are bad for PGN-Random-Books.
Example 1. a3

With random 1. a3 and 1. d4 is the same.

You see, the idea to create such things are good and I can understand what you mean ... but I am sure its not possible to make it perfect (perfect --- your intention I think)

I don't know what is to do now with version 3.5!

Perhaps you have interest to work in the basis file version 3.5 and create an other random-book? I have no problems with it.

What I can do is to analyse the the EPD with 5.120 again with an second engine. Hope the next one will not write ... Frank, we should have three engines for a better selection. I think one engine is enough because Stockfish analysing with a lot of hardware power and time.

Best
Frank

Thinker can do that ... on place 4 after my statistics. Very fine engine! But Thinker don't give any information about evaluations

This is a difficult question. Which engine?
Thinker might be a good compromise really.