Author: Ed Schröder
Date: 05:24:37 12/28/99
Go up one level in this thread
>Posted by Vincent Diepeveen on December 28, 1999 at 06:46:15: > >>You can be right you also can be wrong. When I was 14-15 years I became >>a member of a chess club and left after a few weeks because of all the >>smoke (mainly cigars) and never returned to a chess club. I have played >>several youth tournaments in The Hague (in the AEGON building) and >>from that time is my estimated 1800 elo rating. > >Let's see. 30-40 Years later nothing is left of course of such a rating. >Till around 2200 missing tactics means you lose anyway. This is not the point. The point is if a programmer is still able to understand the moves its creation plays. For that you don't have to be a tactical magician but some good positional skills will surely help. I wasn't talking about winning / losing. And you are wrong about my age too. My age is the elo difference between Rebel9 and Rebel8 estimated at ?? If less even better, you pick... >>>Working on a chess engine sure doesn't improve playing strength, >>>as you let the program solve stuff instead of solving it yourself. >> >>Maybe this is true for you but it isn't for me. I can certainly say my >>positional understanding has improved programming Rebel. My tactical >>skills have gone down a lot because of laziness, lack of interest of >>course due to the fact I have a program for that, why torture my brain >>when a simple mouse click will do the job. > >I must find the first exception to this. Especially someone who 'feels' >he's still 1800 after 30 years. You have a way with words Vincent. >>>If i have worked on diep in the afternoon, then in the evening i play >>>like big shit to be objectively measuring what happens... >>> >>>>Has Rebel improved in playing humans since version 8,9,10 and now Rebel >>>>Century? To answer this question precise you have to realize that hardware >>>>has improved too during the years and people tend not to play old versions >>>>which makes it even more difficult to judge its progress. >>> >>>Very accurate said. Apart from that another aspect needs not >>>to be forgotten: i am used now to fight against crafty at duals >>>and misssilicon at a K6-3, and Hossa at its latest hardware. >>> >>>If i then get against an oldie, i will suddenly do a lot better than >>>i would have done in the past. >>> >>>So where humans have adjusted to the stronger programs, advances in >>>theory, and some other things, the program is still showing the same >>>performance. >> >>If I understand you right you say that "comp-comp" is your only criterion >>to judge Diep's progress? I did the very same in my early days but changed >>that way of testing after 4-5 years. > >No i didn't say that. Biggest eval bugs are found usual because of human-comp. >When humans play for their rating they're deadly accurate in trying to find >a way to beat it :) >Now in contradiction to rebel diep plays a couple of hundreds games each >week against humans. > >However i must add to this that gross errors though they are in all programs, >aren't the only thing you want to solve. > >> >>>>Since times I use the following guide-line to decide which version is best: >>>>- test sets (about 1000 positions) 30% as a first impression. >>> >>>ECM+BK nowadays? >> >>My database is about 85% positional based positions and 15% are about >>tactics. There are about 20-25 ECM positions. Don't know what BK means. > >So you deny having tuned for BK (position 2... d4d5), >and deny knowing wat BK is? I could guess, Bratco-Kopec? If so I don't use it. >Like you never read JICCA, you never read 'computerschaak' and many >other magazines? Hardly. >Hard for me to believe! Then don't. >>>>- auto232 results (30%) >>>>- my personal impression based on my own style and feelings (40%) this >>>>includes the GM challenge games as well. >>> >>>>How do other programmers decide which version is best? and maybe more >>>>important which criteria is involved? >> >>> >>>I test carefully what the evaluation verbosely prints >>>in a position where the bugfixes to the patterns applies to. >>> >>>Positions it played wrong in the past (5000 or something and growing >>>each day nearly, but i only pick a few from which i think apply). >> >>I do the same. And the most worse onces first. >> >> >>>Then it's released to my testers and depending upon their results and my >>>findings i fix bugs in it and decide where to expand again. >> >>You are a reasonable chess player, 2200 I believe. What is your main >>criterion to judge a version? > >What you call 30% feeling is for me actually 100% feeling, >as in the end i'm the judge of everything concerning DIEP. >I must feel happy with how it plays, not the icc folks. Cool. >>>When talking about a non-lineair change of search however i feel it's not >>>so easy to decide. >> >>Right, search solves many positional errors. >>>Let's take for example last ply pruning. It's easy to make last ply pruning >>>such that it does a lot better at testsets. >>> >>>But does it play better then? >> >>Not in my opinion. It just scores better. > >We disagree here. it scores better in blitz. it doesn't play very >well in slow games. You misunderstood. It's what I said too. It scores better in tactical sets but not in games. >>>I find that hard to judge. I have simply thrown all forward pruning >>>out of DIEP and feel a lot more happier. It plays a lot better now, >>>but has a way lower rating in blitz at single cpu machines at icc, >>>the advantage in playing strength can be basically is in my opinion >>>because of evaluation bugfixes. >> >>You are absolutely right about pruning. The main change from Rebel7 to >>Rebel8 was a very narrow selective search resulting in deep ply-depth's. > >I had the impression you also added lazy evaluation to rebel8? Lazy Eval (my own modified version of LE) is in Rebel since times I believe it goes back to Mephisto Polgar (or so). >>To my surprise the thing topped the SSDF list with +65 (or so) while >>Rebel never had been a serious candidate on SSDF. But I also have seen >>the other side of the medal, the holes because of selective search pruning >>essential moves and losing games because of positional blunders. I changed >>selective search a lot since version 8 fixing the holes losing some >>ply-depth but I got back a lot more stable Rebel as a result. >> >> >>>>I also am curious on opinions if Rebel Century is clearly better than let's >>> >>>>say Rebel8 when the subject is playing style which is something different >>>>than playing strength (my opinion and view). >>> >>>I personally feel century is the same engine with a few more tactical >>>extensions and a new book. So i see hardly difference, considering that >>>tactical testsets like ECM, which were solved very bad by rebel8, do >>>not get taken into account in my judgement of engine strength, as i found >>>rebel8 already anything but tactical weak. >> >>Well..... Rebel8 did very well on SSDF. If it was so bad in tactics you can't >>enter the SSDF list as no.1 with +65 on no.2. > >Right when it came out it was tactical the best. Disagree. >However your trick to abort double games with rebel8 >is hugely underestimated worldwide, especially >if you combine that with a book which in some pathetic side lines >which rebel plays well there are a lot of lines int he rebelbook >itself. Making new friends? >>Maybe you can explain to me why ECM is so important for you? Do you use >>all 600/700/800 positions or just a selection? > >It obviously is more important to you, as new rebel versions suddenly >solve ECM positions real soon which rebel8 didn't solve within a >minute or 10 at least. Dieps behaviour on ECM hardly has changed. ECM is not important for me as it is only about tactics. I have enough of these. Ed > >Vincent
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.