blogPosted by Peter Fendrich Friday, July 06 2007 13:39:13
I am in the process of publishing a new version.
However I first want to do 3 things:
- There is a book code bug that I expect to solve.
- Node count tests between 2 test versions that should have the same node counts
- Mirror tests. The same position as white and black should give the same results.
I have a new book from Günther Simon that seems to be good. I have only made a few shallow tests but sofar it looks good to me. Thanks Günther!
blogPosted by Peter Fendrich Monday, June 25 2007 00:39:08
I am using Baron, Ruffian, WildCat7 and Smarthink as sparring partners.
My own tests give these results:
5min+3sec: About 80 ELO better than previous version (704)
40min/40moves: About 150 ELO better than previous version (704)
I also got 1min+1sec tests from another tester and he found this version about 40 ELO better than previous version.
It could be interpreted as the deeper the bigger is the difference...
but we talk about 100 game tourneys and some error margins so I will wait with too detailed statements.
I will do some more tests and also compare with other testers results before releasing a new version.
blogPosted by Peter Fendrich Tuesday, June 19 2007 16:18:06
I have fixed a serious bug in the search and after testing I got some amazing results.
I know that I am always far too optimistic and should not rely on my own testing
I need som really objective testers!
I hope also to track down that bug I have lived with for 2 years or more...
blogPosted by Peter Fendrich Sunday, June 17 2007 23:44:22
I had more problems than I expected using my method to get minimum sized bitbases.
KRPKR was quite complicated to get right and I will take a pause and concentrate on other things for a while. I will look at some pruning ideas and increased king safety.
The bitbase idea have to get some adjustments tha I will think of at the beach...
blogPosted by Peter Fendrich Monday, May 28 2007 21:21:54
...and the time flies.
The KRPKR just takes time to get ready. I am currently running one of the shots that takes about 8 hours. I had several bugs and everyone took at least one day to solve. Sofar this run looks ok. If this one doesn't work out I will try another approach with traditional bitbases for part of the solution space.
The question is how I can get ready before summer.
New report will follow soon.
blogPosted by Peter Fendrich Friday, May 11 2007 17:27:14
Currently Alaric has only one bitbase (a complete database of an endgame). The bitbase kpk.dat that you you will get when you download Alaric covers the endgame with King and pawn versus King. It contains one bit for each possible position with '0' for draw and '1' for win - the only 2 results possible in this endgame. That gives about 120000 positions when I use mirroring and some other tricks to keep the size down. Gives about 44 Kb.
I tested a new technique where I combine chess knowledge and trees for the same endgame. Now I need only 4 Kb and it is complete, that is all possible positions are covered by these 4 Kb. I will test also K, R, P vs K, R to see what can be done. If that works out I will complete all 3- and 4-men endgames with the same technique.
5-men? Well let's see how this works out first...
It is quite a manual task to identify what kind of knowledge to use and how to represent it.
My first intention to write a bitbase generator and compress the bitbases as much as possible is somewhat changed. Look here.
blogPosted by Peter Fendrich Saturday, March 31 2007 18:06:42
Most of the black moves were missing in the book I uploaded in DannsBook.zip
It is now corrected and I call it DannBook.zip just to get a new name for the new version...
blogPosted by Peter Fendrich Tuesday, March 27 2007 22:18:17
The new version will be out in a day or so.
The version is 703. I used to have an extra zero in the version name like 0703. From now on I will remove that. 703 means 2007 mars...
I think that my new reduction functions will pay off in strength.
I hope that the tournyes out there will confirm this
A new book learning scheme is developend but it is turned off in this version.
I will test it out better before releasing it.
Future versions will have more extensions, especially will I try to detect patterns for king attacks and maybe other types of attacks and threats.
My sparring partners these days are:
They represent a good variety of playing styles. Thank you Lars!
I check out my changes against them and will not update their version numbers even if they are upgraded. I need a stable test bench!
Maybe one day I even dare to include Glaurung in the future but that engine is still to strong for me and will maybe be forever...
Dann created a new book for Alaric with some secret weapons included.