Author: Robert Hyatt
Date: 09:57:30 01/14/04
Go up one level in this thread
On January 14, 2004 at 11:23:54, Reinhard Scharnagl wrote: >On January 14, 2004 at 11:10:55, Tony Werten wrote: > >>On January 14, 2004 at 10:47:56, Reinhard Scharnagl wrote: >> >>>On January 14, 2004 at 10:35:38, Uri Blass wrote: >>> >>>>On January 14, 2004 at 10:19:03, Reinhard Scharnagl wrote: >>>> >>>>>On January 14, 2004 at 07:52:19, Uri Blass wrote: >>>>> >>>>>>I think that learning can be very effective. >>>>>> >>>>>>An engine that does not learn may lose the same games again and again after >>>>>>enough games. >>>>>> >>>>>>I use learning for matches of 4 games that are popular in Leo's tournament and >>>>>>my learning is simply to choose a different first move after a loss. >>>>> >>>>>Is this real 'learning', or an escaping into a not yet refuted randomizing? >>>>> >>>>>>With my very small manually edited book(only few hundreds of positions) there >>>>>>are big chances that movei will lose the same game twice if I do not do it. >>>>>> >>>>>>For testing I prefer to use the nunn2 match and test suites. >>>>> >>>>>Nevertheless that behaviour really may produce success, it is not what I would >>>>>call learning. But you are not alone using the word 'learning' that way. >>>>> >>>>>Before claiming something being able to learn, please specify, what is learning. >>>>>I still cannot do this sufficiently. >>>>> >>>>>Regards, Reinhard. >>>> >>>>Every behaviour of a program that is dependent on the history of games is >>>>learning. >>> >>>Hello Uri, >>> >>>learning is possible from success or from failures. (And I hope not to have >>>made you angry by the above.) >>> >>>Failures (in opposit to successes) mostly can be localized at a special point of >>>history (you correctly demands that dependance). >>> >>>But loosing a game can be completely independent from the opening moves. >>> >>>Without being able to localize the probably point of error (with a lot more than >>>low random chance) how could there be a correct implementing of experiences? > >>Does it matter ? > >Yes. Doing things like that avoids 'real' learning. > >>If opening theorie says a certain book position is winning, yet >>your engine keeps loosing it, you'd better avoid that position. > >May be. As I have written, it might be successfull, especially within badly >minimaxed opening books. > >Would you decide to never leave your home, when something evil has happened to >you outside? When we all would practice such a kind of 'learning', we all have >to commit suicide - seems to be a bad idea. I think that is an extreme case, more like choosing to totally turn the book off after one loss. In real life, I might choose to avoid the place where something bad happened, yes. But that doesn't mean that I don't leave my house at all. > >>Of coarse, if you change your engine and suddenly it does understand the >>position, you'd better throw away the learnfiles. > >Of course, but you still should try to learn by avoiding errors. I believe that is what reasonable book learning does... You might visit my web site, which now has links to a half-dozen papers I have written including the one on book-learning... it explains what I do, at least, and how effective it can be. www.cis.uab.edu/info/faculty/hyatt.html if I recalled that correctly. > >Regards, Reinhard.
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.