Author: Roberto Waldteufel
Date: 02:02:19 06/20/98
Go up one level in this thread
On June 20, 1998 at 01:28:44, Don Dailey wrote: >On June 19, 1998 at 23:40:17, Roberto Waldteufel wrote: > >> >>On June 19, 1998 at 15:55:13, Steven J. Edwards wrote: >> >>>On June 19, 1998 at 15:04:21, Tim Mirabile wrote: >>> >>>>I've never been a big fan of this method. Even high level games tend to have >>>>results which do not necessarily indicate who played best in the opening. But >>>>it's hard to suggest another method that does not involve a lot of hand tuning. >>>>Perhaps you could just suck in all of ECO and Informant, with evaluations, which >>>>if not totally trustworthy would be more so than game results. >>> >>>I dislike the uniform distribution method mostly because it seems that a lot of >>>information is being thrown away. But I also dislike the ECO/Informator >>>transcription (I hear these can be had on CD-ROM nowadays) because it is >>>particularly vunerable to attack by typo hunters who peruse the source input. >>>Additionally, the Informator evaluations are really not scalars and some sort of >>>multivariate weighting function would have to be used. And so we're back to the >>>same type of problem as with weighting win/draw/loss results. >>> >>>I can think of ways to have a program tune its own book from PGN input files, >>>but they all involve playing through the lines with heavy duty analysis to >>>locate variations which conform to the program's search and evaluation. >>> >>>One idea I've considered is getting a couple of books with titles like _White to >>>Play and Win, a Complete Opening Strategy for the Attacking Player_ and just >>>copying the variations into the program's book. I think that a 5,000 move book >>>would suffice for this, but once the opposition figured out what was going on, >>>then it would be time for another book. >>> >>>-- Steven (sje@mv.mv.com) >> >>Hi Steven, >> >>I have often pondered this question, and like you I have failed to find a >>satisfactory answer. In the end, I don't think that mere vollume is an adequate >>replacement for chess judgement. Probably the most effective way to build an >>openings book is to do it by hand, but this is, as you know, very slow and >>tedious. Your suggestion about using opening repertoire publications is quite >>similar to my approach, but there is something that I might suggest you try in >>order to help in your selection. The openings should be chosen with great care >>to compliment the strengths and weaknesses of your program. Ideally you want >>openings that, although they may not confer an objective advantage, lead to >>positions that your program handles well. If your program is best in very >>tactical positions, you might consider giving it some sharp gambit lines to play >>for example. When you have just entered a new line of play, say for White, try >>letting the program play against itself. Does White win? If so, your choice was >>well founded. If White loses, you either have to supply more variations to "plug >>the knowlege gap", or else you scrap the whole line and substitute something >>else instead. In this way, you can slowly build an opening book that is tailored >>to your particular program, which I think is also similar to the way that a >>strong human player chooses an opening repertoire. I hope you find the idea >>useful, >> >>Roberto > >Hi Roberto, > >I think your idea has much merit. I've always advocated picking lines >that suit the programs style. Playing the program against itself >might give some clues about how good the choice is. > >I would like to suggest another approach, one that I rarely use but >I think also has merit. We all tend to constantly adjust the >book and much has been posted on this subject from Bob and others >including myself. But maybe we should consider the possiblility >of adjusting the program to the book instead (or in addition to) >adjusting the book to the program? There have been times when >Cilkchess comes right out of book and quickly makes a positional >error or weak move. I have sometimes fixed the cause of the >error since I know there is a problem. I usually then also >fix the book just to be on the safe side. In principle we >have two "knobs to turn" not just one. > >If it's a more general case of the program just not understanding >the opening ideas then I still think it might be an opportunity >to improve your program instead of just ignoring the problem by >tunning the book. > >I believe the hardest program to beat would be the one with a >big wide opening book where the program plays each system >reasonably well. Building one of these of course is no easy >task! > >- Don Hi Don, Yes, this is something I never thought of doing. In my case it poses a curious problem in that my program has a very different style of play to my own. Many of the openings it uses are openings that I know comparatively little about, since I never meet them in my own games (I always open 1.d4 with White, but my program likes 1.e4 openings best). Since the program generally plays a bit better than I do, and I don't always feel qualified to say what the best plan in an opening is, so I have to rely on advice from publications and/or stronger players who do use the openings in question, but then they do not always agree. I like your idea in principle: it should enable the program to handle more and more openings correctly, but I think there might be cases when you make a change to the program that succeeds in improving its play in a particular opening variation, while maybe having an adverse affect on its play in another situation, and, what is worse, with a really big opening book it would be impractical to test the change in all the lines, or even in a large selection of them. I think this problem would make it quite difficult to actually know how well the program tuning was working. The advantage of altering a single opening variation rather than a single evaluation term in the program is that you know that nothing else can change as a result, so you can more easily test whether the change has helped or hindered the program. Best wishes, Roberto
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.