Author: Mike S.
Date: 12:02:24 10/18/00
Go up one level in this thread
On October 18, 2000 at 14:25:33, Uri Blass wrote: >>(...) >>1. switch off all (permanent) learning >>2. remove all doubles (!!) from the results > >In this case you may get a match of only 2 games when both sides always repeat. I think any book has at least some variety even at tournament setting. Few exceptions. >(...) >Suppose a program learns to have better evaluation function from games(I do not >know about programs who do it but I believe that it is possible). Yes, it was said that Deep Blue did this, and one of the older Nimzo versions (3 I think) has a function too, to distillate learning data from game databases. But this were additional features, not an automatical learning from the games the program played. Nevertheless, all those learning functions are interesting of course, but I think they should be examined in special tests, comparing the effectiveness etc. I can imagine that such a test would be difficult to design. The documentation of learning functions is also not satisfying IMO. For example the question of different time controls, or different hardware speed... What the program thinks to be a mistake on P200, may look differently if I upgrade to PIII/800. Does the program use the data as if it was produced on the faster computer, or not? The same question applies for different time controls. As a customer, I never know exactly when and how a program learns, and when and how this data is used. There is a lack of in-depth manuals accompanying the programs. Regards, M.Scheidl
This page took 0 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.