Author: Dave Gomboc
Date: 11:11:39 10/09/99
Go up one level in this thread
On October 09, 1999 at 13:55:46, Ratko V Tomic wrote: >That seems to be dependent on how frequent given opening line was in its book. >If it was rare with, say, just a handful of wins for one side, even a handful of >games makes the significant change. In my case, after it locked into the same >line (nimzowitsch opening), the book showed 100% vs 0% for one side, and no >matter how you scale that it still remains 100% vs 0%, i.e. the most successful >line. At that point I had an option to check the literature and find a good >counter and then win enough games in that line to offset the nonscalable ratio >or reinstall the book from scratch. Reinstalling seemed quicker, so I did that, >and set the learning influence to minimal, variety to maximal, but few days >later again it locked into the single line (again nimzo). So at that point I >reinstalled the book again and turned off the learning. But a day later when it >started again playing nimzo several times in a row, I finally reinstalled the >book and turned the file attributes to read only, and that finally made the >variety of openings Ok. Another problem with "learning" was that at some point >it threw away entire Sicilian from its repertoire after losing couple games for >reasons well beyond the opening. I am not sure why, their bugs and poorly >thought out design choices are sometimes hard to distinguish (they did have a >good artist, their stuff is visually the best of the bunch). Wiping e.g. the Sicilian from its repertoire is probably a reaction to "that other computer seems to play these kinds of middlegames better than I do. I'd better find an opening system that it doesn't understand so well." Of course, if it's completely overmatched, it should retain variety because everything will look bad. :-) I don't disagree at all that this behavior is extremely undesirable from a sparring partner standpoint. Dave
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.