Author: Ed Schröder
Date: 15:03:15 03/18/98
Go up one level in this thread
On March 18, 1998 at 13:13:34, Bruce Moreland wrote: > >On March 18, 1998 at 05:41:06, Ed Schröder wrote: > >>This is no joke. I checked the Rebel9 book learn software and I can >>do it if I wish. This is crazy no? It's crazy because the 30-40 elo >>improvement is counted as a gain of playing strength. That's what >>the list implies or? >> >>And now the $64,000 question.... >> >>Shall I? >> >>It's a cheat no? >> >>It's a cheat because it hides the REAL strength of an chess engine. > >Make a great learner. I'll never accuse you of cheating. You can >promote it as the newest greatest feature in Rebel 10 and you won't hear >a peep from me. > >Learners are great. It's an area of computer chess that has been >neglected but now has been brought back into play because of you guys >competing with each other (trying to dick each other) on the Swedish >list, and to a much lesser extent because of the success of Crafty's >learner on ICC. > >A learner is as honest and useful as null-move search or what the >germans call "permanent brain". > >If you don't have a learner, your domain is the single game. Chess >programmers have been competing in repeated instances of single games >for years and years. But this has always been a flaw in the computer >approach. > >This "REAL" strength you speak of is incomplete and sadly limited. >Innovation that goes beyond this is not cheating. > >The human approach is to try to improve your play, or at least avoid >playing the same losing line N times in a row. > >It is an *advance* in the state of computer play when computers do the >same thing. > >With a learner, your domain is the match. You no longer have this >utterly unrealistic spell of amnesia after every game, you can adapt to >what your opponents hit you with. > >Add a learner that can handle with interrupted matches. Furthermore, >feel free to enhance it so it understands the traits of more than one >opponent (not necessarily so that it knows that it is playing "Genius", >but rather that it is playing "Bob" instead of "Fred"). This allows a >program to go beyond the match domain, and enter the career domain. > >A related issue involves how the Swedish list people do testing. >Perhaps the conditions under which these programs compete need to be >more fully specified. If there is a single entry for this wonderful new >learning (if I convince you) "Rebel 10", there should be a single entity >called "Rebel 10", and that entity should be able to benefit from >experience gained by every game this entity plays. > >I don't know if it would be practical to do this, but the Swedish list >folks need to take extra pains to run these matches fairly when there >are programs that are operating within the domain of the match or career >instead of the single game. > >And perhaps the programmers need to be involved by helping to specify >the conditions under which their programs compete, and possibly by >creating utilities that make the Swedish testers' jobs easier (for >instance, by creating something that allows you to export your learning >information, so one tester can email it to another tester who wants to >run a match with your program). > >bruce Thanks Bruce for your clear view on (book) learning. So everything is allowed? So maybe we should look in a different way to the SSDF. Quite confusing for people I fear. Introduce a new second "error margin"? An error margin for learning? :)) - Ed -
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.