Author: Moritz Berger
Date: 02:54:46 09/10/98
Go up one level in this thread
On September 09, 1998 at 20:15:12, Don Dailey wrote: >On September 09, 1998 at 11:34:10, Moritz Berger wrote: >>The Fritz PowerBook is compiled from a GM database with the only criterion for >>including a game in the book being the playing strength of both players. >> >>The PowerBook doesn't contain *any* preferences to make Fritz prefer or avoid >>any kind of opening lines, except of course the statistics from the human >>database (win/loss/draw ratio and move frequency as well as ELO performance >>which Fritz doesn't use AFAIK). It's basically just a good database, nothing >>more and doesn't include any previous Fritz games. >> >>Moritz > >Moritz, > >I was under the impression that they used a learning process of some >kind. I got this directly from Franz although he provided little >detail. He was disapointed after getting a bad result at some >tournament and thought perhaps the idea was no good. I also got >the impression he was not directly involved in the programming of >the book learning method but I don't know this for sure. Fritz does learning, of course. But the PowerBook doesn't have any pre-learned parameters for the Fritz engine, it learns as the matches against other opponents (no matter if human or silicon) go along. Implication: Fritz plays all 1.e4 1.d4 1.c4 1.Nf3, all popular responses to 1.e4 e.g. e5, c5 ,e6 you-name-it. As long as a particular opening has enough followers among IM and GM players, Fritz will play this, too. One weakness in Fritz' approach is that it even follows games with only 1 occurence in the database up to the very last book move, unlike Crafty where you can specify e.g move frequency >2 to blindly follow a path. >But soon after Franz telling me about this I notice Fritz shoot >to the top of the rating list. In my mind I suspected this was >related to the book work I knew they had done but it was only a >guess and I have no way to know. It could be a completely >unrelated event for all I know. The book learning works quite well at the SSDF, but - unlike anybody else - Fritz doesn't even have a start.PGN file like Crafty to influence the most basic preferences, i.e. 1.d4 over c4 and so on. It gets into closed positions and positions it doesn't understand too well (from a human point of view) quite often, but at the SSDF it only plays other programs that face similar "computer" problems with these positions so it doesn't get punished as much as it would against humans. A good example is the fact that Fritz plays 1.c4 (and gets punished against several opponents for doing so), but still scores the usual ~70% at the SSDF as far as I remember ... >So I am curious about where you got this information. Is this common >knowledge, part of the chessbase documentation or did you learn >this from some other source? Matthias Wüllenweber described somewhere in detail how he built the PowerBook (he named the databases and ELO strength selection criteria he used). Fritz shows book weightings for all move in the "tree" view, they are all 0 by default, the resulting move probabilities (which are also shown) correspond to the underlying database results (that are also visible on the same line). But of course I don't trust ChessBase and have built many own books, ranging from 400 MB to 2.7GB in size from various databases of different quality. Results didn't differ much from the results with Fritz' PowerBook. Fritz even got 50% in >30 tournament games with a book I built from 1000 Anand games against another TOP 5 program... Conclusion: Fritz has demonstrated that opening books don't matter enough to keep a program away from the #1 at the SSDF, although it might have scored even better with a customized book. > I am pretty interested in knowing this >myself and you seem to be in the know here. From what you are saying I >am starting to believe they have a completely separate in-house book >they use and Fritz got it's impressive results without any special >book work or tuning of any kind. Chessbase did use a tournament book in the Frankfurt rapid chess event, but this book is not available to the SSDF. Of course, using a custom tailored book is even better than playing with a database dump ... Moritz
This page took 0.01 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.