Hi,
Manhattan.abk has been uploaded at WBEC, into Romichess page, here:
http://wbec-ridderkerk.nl/html/details/RomiChess.html
Manhattan book is in Arena format.
For those who don't know about my recent tests, I run tests The Baron 2.3-Crafty 20.14 (300 games), The Baron2.3-Glaurung 1.2.1 (1100 games), and Romichess P3i-Crafty 20.14 (750 games).
Hardware was AMD Athlon 4000+ with 1 GB RAM.
Time controls were 1' for game +1" each move.
In first test Baron opponent was Crafty 20.14 using its opening book. Here are Baron performances:
Games 1-100 +37 -31 =32 (53%)
Games 101-200 +40 -32 =28 (54%)
Games 201-300 +49 -26 =25 (61,5%)
Second test opponent was Glaurung 1.2.1 with his opening book. Here are Baron performances:
Games 1-100 +11 -63 =26 (24%)
Games 101-200 +17 -61 =22 (28%)
Games 201-300 +12 -65 =23 (23,5%)
Games 301-400 +27 -53 =20 (37%)
Games 401-500 +23 -56 =21 (33,5%)
Games 501-600 +22 -53 =25 (34,5%)
Games 601-700 +23 -48 =29 (37,5%)
Games 701-800 +19 -49 =32 (35%)
Games 801-900 +20 -52 =28 (34%)
Games 901-1000 +24 -54 =22 (35%)
Games 1001-1100 +28 -39 =33 (44,5%)
Third test was Romichess P3i against Crafty 20.14 and its opening book. Here are Romichess performances.
Games 1-100 +16 -74 =10 (21%)
Games 101-200 +32 -49 =19 (41,5%)
Games 201-300 +28 -59 =13 (34,5%)
Games 301-400 +33 -46 =21 (43,5%)
Games 401-500 +37 -56 =7 (40,5%)
Games 501-600 +46 -38 =16 (54%)
Games 601-700 +48 -35 =17 (56,5%)
Games 700-750 +25 -16 =9 (59%)
Manhattan.abk book has a secret. It has only one move for learning engine side, but it has all main opponent replies. This way, number of variations to learn about are very limited (less than 1000). Thanks to traspositions, total variations could be not more than 600.
This structure helps engine with effective learning systems, like Romichess and The Baron. Learning obstacle has always been the huge number of variations from starting position. With Manhattan book I managed to solve this problem.
Future plans include a larger book building with the most popular openings included. But that work will take several months.
I hope somebody will be interested to run some tests, and to improve book itself. And I'd like to hear impressions and opinions.
Bye,