Finally....
Moving away from windows and into Linux.
Installing Linux and getting things running to a looooong time, changing the interrupt code in Jabba took about 2mins... otherwise, it compiled ok, and with GCC 4 is at least as fast as MSVC++ Express on windows.
If there are any bugs, please let me know.
Sunday, February 21, 2010
Sunday, January 24, 2010
More work on Jabba
Work is continuing on Jabba.
First, during completion of the cStats class for analysing game databases, I ran a test on 900,000 games to make sure Jabba could process every move, as a way to test the pgn parsing -> namely the SAN conversion functions. I found quite a few bugs, which has caused some changes in the bookmaking code.
The bookmaking code has also been refined, collecting win percentages for white and black, and various other move statistics.
I'll be a able to build what should be a much stronger book now, mainly due to the fact that I'm using 300,000 Playchess.com machine room games for it! :)
Other work includes the evaluation. Psqt's have been replaced, bishop mobility added, and opposite side castling psqt have also been added. I've also added quite a bit of endgame knowledge, from various chess books I have, regarding special drawn / won positions for KPvsK, BvsP, NvsP, RvsP and a few others. Unfortunately, I got rather carried away with writing them and not testing them, and there are many many bugs which will take a few weeks to iron out!
Then, I need to test each feature through the STS suite, to see how the changes have helped, if at all....
First, during completion of the cStats class for analysing game databases, I ran a test on 900,000 games to make sure Jabba could process every move, as a way to test the pgn parsing -> namely the SAN conversion functions. I found quite a few bugs, which has caused some changes in the bookmaking code.
The bookmaking code has also been refined, collecting win percentages for white and black, and various other move statistics.
I'll be a able to build what should be a much stronger book now, mainly due to the fact that I'm using 300,000 Playchess.com machine room games for it! :)
Other work includes the evaluation. Psqt's have been replaced, bishop mobility added, and opposite side castling psqt have also been added. I've also added quite a bit of endgame knowledge, from various chess books I have, regarding special drawn / won positions for KPvsK, BvsP, NvsP, RvsP and a few others. Unfortunately, I got rather carried away with writing them and not testing them, and there are many many bugs which will take a few weeks to iron out!
Then, I need to test each feature through the STS suite, to see how the changes have helped, if at all....
New Tournaments & Results
Jabba has finished it's tournament "Battle of the forgotten", finishing in a respectable 3rd place, although in a weak field.
Jabba has also (Very kindly from the director) been included in the 17th Edition of WBEC tournament, where Jabba is playing in the 5th Division Group C. I'm not expecting Jabba to qualify for promotion - Lime only just managed that, and Jabba is 200 elo weaker, at least. But, it will give me the first set of long time control games to analyse.
Another tournament has finished, HG Mueller's Knightmate Challenge, where Jabba, under time odds, managed 5th from 11.
The tournament also demonstrated a problem with the plain Alpha-Beta search... Jabba saw a win, involving promotion, in one game, but never actually pushed for the winning line, just holding the promotion in the pv until it was too late, and the game was drawn by 50 move rule. See the above link also for the game!
Jabba has also (Very kindly from the director) been included in the 17th Edition of WBEC tournament, where Jabba is playing in the 5th Division Group C. I'm not expecting Jabba to qualify for promotion - Lime only just managed that, and Jabba is 200 elo weaker, at least. But, it will give me the first set of long time control games to analyse.
Another tournament has finished, HG Mueller's Knightmate Challenge, where Jabba, under time odds, managed 5th from 11.
The tournament also demonstrated a problem with the plain Alpha-Beta search... Jabba saw a win, involving promotion, in one game, but never actually pushed for the winning line, just holding the promotion in the pv until it was too late, and the game was drawn by 50 move rule. See the above link also for the game!
Sunday, January 17, 2010
Jabba 1.1 Mac
Last night, I compiled Jabba on the Mac... called version 1.1, because it has some changes (in progress) from version 1.0. Strength is pretty much identical.
Windows version will stay at version 1.0 until the next "proper release".
Link on the right.
Windows version will stay at version 1.0 until the next "proper release".
Link on the right.
Working on Jabba
I'm working quite hard on Jabba's eval. The released version has almost no eval.
For tuning, I'm using the STS suite, rather than games.
I've finished writing a new class, called cStatmake, which analysis and produces statistics from game databases.
So far, I've collected material balance stats, using a database of 1 million Playchess engine room games. I have then used regression in MS Excel to obtain values for BPair by no. of pawns and majors, for BvsN and for the exchange.
Since then, the piecce square tables have all been replaced, bishop mobility has been added, and knight support points. Next step is to add some endgame knowledge, and to use the database to assess passed pawn values.
Once the STS suite score is improved by a good margin, I'll look at adding SEE to prune captures in Q Search, which should bring a nice speed up.
The aim is to have something ready for the CPT 2010, in March.
For tuning, I'm using the STS suite, rather than games.
I've finished writing a new class, called cStatmake, which analysis and produces statistics from game databases.
So far, I've collected material balance stats, using a database of 1 million Playchess engine room games. I have then used regression in MS Excel to obtain values for BPair by no. of pawns and majors, for BvsN and for the exchange.
Since then, the piecce square tables have all been replaced, bishop mobility has been added, and knight support points. Next step is to add some endgame knowledge, and to use the database to assess passed pawn values.
Once the STS suite score is improved by a good margin, I'll look at adding SEE to prune captures in Q Search, which should bring a nice speed up.
The aim is to have something ready for the CPT 2010, in March.
Jabba's first tests
So, jabba is now playing in some tests.
First, "Battle Of The Forgotten", a tournament by Carlos Pagador.
Results are here.
A tournament for weak engines, but a. Jabba's looking good for the win, and b. no reported bugs, which is the most important thing.
Second up is "CET2403y" by Harun Taner, who is using both the 32 and 64bit versions. The result is also ok here, and most importantly, no bugs reported. Link to tables
First, "Battle Of The Forgotten", a tournament by Carlos Pagador.
Results are here.
A tournament for weak engines, but a. Jabba's looking good for the win, and b. no reported bugs, which is the most important thing.
Second up is "CET2403y" by Harun Taner, who is using both the 32 and 64bit versions. The result is also ok here, and most importantly, no bugs reported. Link to tables
Saturday, January 16, 2010
Jabba 1.0 Knightmate Variant
In response to a new tournament by HG Mueller, posted on Talkchess.com, I've made a version of Jabba that plays the Knightmate variant of chess. See download link on the right.
Subscribe to:
Posts (Atom)