Posts : 2571 Join date : 2020-11-17 Location : Netherlands
Subject: ProDeo 3.0 Tue Nov 17, 2020 9:34 am
Playing strength first, I collected a number of changes during the last 4 years resulting in a ~35 elo improvement. ~20 elo came from a new idea replacing IID with what I called IIR. A number of other engines (among them Stockfish) adopted the idea.
Polyglot opening books, improved the standard ProDeo opening book with +46 elo, see elsewhere. Created additional books with new options, an overview:
But the book ideas are pretty cool. This comes +extremely+ close to what Marc Lacrosse and me did in 2004/2005, +manually+ ( investing an +insane+ amount of time) I may add .
It is cool to see what a programmer can do with such a setup ( minus some of the things book cooks can still do better). I think you default to a WAY too long book depth btw, but I really like the general setup.
Mclane
Posts : 3011 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
I wish Ed you would stop testing against older Rebell versions. IMO this only measures how different a new version is. The opponents have to be 4-8 different strength and style engines.
Only One opponent is IMO giving wrong data,
Admin Admin
Posts : 2571 Join date : 2020-11-17 Location : Netherlands
Subject: Re: ProDeo 3.0 Fri Nov 20, 2020 8:41 pm
Peter Berger wrote:
ProDeo is still ProDeo/Rebel.
But the book ideas are pretty cool. This comes +extremely+ close to what Marc Lacrosse and me did in 2004/2005, +manually+ ( investing an +insane+ amount of time) I may add .
It is cool to see what a programmer can do with such a setup ( minus some of the things book cooks can still do better). I think you default to a WAY too long book depth btw, but I really like the general setup.
You remind me I still have to test the new book for blunders, forgot all about it due to the setup of the forum. Although the book is made from elo 2700+ humans players it's not unlikely it contains a number of losing variations. I have started the procedure. Basically, run 2000 games at a decent time control with SF12 and modify the book with the POLY tool. Nowadays everything goes automatic, no more hand tuning.
Mclane likes this post
Admin Admin
Posts : 2571 Join date : 2020-11-17 Location : Netherlands
I wish Ed you would stop testing against older Rebell versions. IMO this only measures how different a new version is. The opponents have to be 4-8 different strength and style engines.
Only One opponent is IMO giving wrong data,
Stockfish exclusively are doing incest testing, seems to work quite well. Besides, I am not very interested in playing strength any longer, just following the progress others make is fascinating enough. Playing style is a different issue, Benjamin 1.0 was a good start.
Mclane likes this post
Mclane
Posts : 3011 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: ProDeo 3.0 Fri Nov 20, 2020 9:14 pm
Yes you are right, Benjamin 1 was ok. I liked the style it did.
Ever thought about making prodeo SMP ? Today any person has a pc with 2-64 cores, even my mobile phone has 8 cores.
Admin Admin
Posts : 2571 Join date : 2020-11-17 Location : Netherlands
Subject: Re: ProDeo 3.0 Fri Nov 20, 2020 9:34 pm
There was a time I had to do things I disliked, things that had nothing to do with chess, programming the leds on the board, programming the keys, programming the clock, creating your own character set for the screen, scanning the board for a move. Programming SMP is just like that, nothing to do with chess, I am free now.
Mclane and matejst like this post
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: ProDeo 3.0 Fri Nov 20, 2020 9:42 pm
Admin wrote:
There was a time I had to do things I disliked, things that had nothing to do with chess, programming the leds on the board, programming the keys, programming the clock, creating your own character set for the screen, scanning the board for a move. Programming SMP is just like that, nothing to do with chess, I am free now.
That’s what I thought, but actually SMP is just a few lines of code and a header file threads.h or pthreads.h The basic idea is well known, very simple to implement. And assuming one’s engine isn’t writing to globals all over the place.
Mclane likes this post
Admin Admin
Posts : 2571 Join date : 2020-11-17 Location : Netherlands
Subject: Re: ProDeo 3.0 Sat Nov 21, 2020 10:54 am
It does have globals all over the place, secondly I need to switch compilers, 1+1=NO.
Peter Berger
Posts : 130 Join date : 2020-11-20
Subject: Re: ProDeo 3.0 Sat Nov 21, 2020 7:20 pm
Admin wrote:
Peter Berger wrote:
ProDeo is still ProDeo/Rebel.
But the book ideas are pretty cool. This comes +extremely+ close to what Marc Lacrosse and me did in 2004/2005, +manually+ ( investing an +insane+ amount of time) I may add .
It is cool to see what a programmer can do with such a setup ( minus some of the things book cooks can still do better). I think you default to a WAY too long book depth btw, but I really like the general setup.
You remind me I still have to test the new book for blunders, forgot all about it due to the setup of the forum. Although the book is made from elo 2700+ humans players it's not unlikely it contains a number of losing variations. I have started the procedure. Basically, run 2000 games at a decent time control with SF12 and modify the book with the POLY tool. Nowadays everything goes automatic, no more hand tuning.
Yes, blunderchecking the book is worthwhile IMHO, and Marc and me did this for hours on end.
All these lines were entered into the book the day before the game btw. The Crafty version in Turino was actually +way+ weaker than the one that played in Ramat-Gan ( programmers didn't test like they do now) , but once THIS line was on the board I knew Diep was toast ( and Vincent knew, too ) - those were the days . Last book move was 17.. a5 and Crafty still had to find Re8 on its own , I sweated a little
Admin and TheSelfImprover like this post
Admin Admin
Posts : 2571 Join date : 2020-11-17 Location : Netherlands
Yes, blunderchecking the book is worthwhile IMHO, and Marc and me did this for hours on end.
Totally. But I think it would also be cool to keep the original (or an even weakened one) for human play.
Perhaps in UCI options can choose "Strong book" and "Human book" from a dropdown or something?
It is quite common in human play to face people playing slightly inferior lines, but not very common to face people whose entire repertoire is on Caruana's level.
Really like this book version. Even going through lines in analysis mode is cool looking at the info window and seeing that Stockfish NNUE played 13.Rc1 here recently and Carlsen played 13.b3 last year and so on.
Very nice work.
Mclane
Posts : 3011 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: ProDeo 3.0 Sun Nov 22, 2020 1:21 pm
As I said before, the results of testing Version x+1 against Version x is not making much sense regarding ELO. The ELO works only if the opponent is the old engine. But in the moment other engines enter the battle, the ELO has no base anymore.
Therefore when I test I do usually test against many different engines. Strong ones, weak ones, old and new ones. Stupid ones , intelligent ones.
The overall ELO is then the real one.
When programmers Mainly test against old versions of their own engine, the reason is often that their testing system can produce large amounts of games in a very fast time. They are used to testing this way. And they eliminated errors. But the question still remains: is the „progress“ you measure similar if the new engine plays versus others. And here I don’t think the Stockfish method works.
supersharp77
Posts : 26 Join date : 2020-11-25
Subject: Re: ProDeo 3.0 Wed Nov 25, 2020 10:32 pm
Admin wrote:
Playing strength first, I collected a number of changes during the last 4 years resulting in a ~35 elo improvement. ~20 elo came from a new idea replacing IID with what I called IIR. A number of other engines (among them Stockfish) adopted the idea.
Polyglot opening books, improved the standard ProDeo opening book with +46 elo, see elsewhere. Created additional books with new options, an overview:
Well I went to the Rebel site and was going to download BUT...This engine was HUGE....1/2 Gig or some huge number....I would download just the engine only and the books and personalities..how come its so big? Thx AR....
Admin Admin
Posts : 2571 Join date : 2020-11-17 Location : Netherlands
Subject: Re: ProDeo 3.0 Thu Nov 26, 2020 1:48 am
supersharp77 wrote:
Admin wrote:
Playing strength first, I collected a number of changes during the last 4 years resulting in a ~35 elo improvement. ~20 elo came from a new idea replacing IID with what I called IIR. A number of other engines (among them Stockfish) adopted the idea.
Polyglot opening books, improved the standard ProDeo opening book with +46 elo, see elsewhere. Created additional books with new options, an overview:
Well I went to the Rebel site and was going to download BUT...This engine was HUGE....1/2 Gig or some huge number....I would download just the engine only and the books and personalities..how come its so big? Thx AR....
It's called the "Book Edition". The focus is on opening books and some of them are big.
Download size is 290 Mb, is that really a problem nowadays?