Author: Ed Schröder
Date: 00:58:38 08/10/99
Go up one level in this thread
>Posted by Eelco de Groot on August 09, 1999 at 16:05:35: > >Thanks for trying that out, Paulo! >That would be just about what you would expect, I suppose, if it follows the >rule of thumb Robert Hyatt posted the other day; that the best move takes up >about half the time of a normal search and that for n-best search for n>1 you >can expect to have to add n-1 times that amount. In this case with n=6 it >would be more or less 5X150=750 seconds to add to 300. But the key move >wasn't found (I certainly do hope it was somewhere near the first six?) so this can >mean for >'really difficult' positions you have to search at least one ply deeper to get >the same level of accuracy in determining the best move. > >Not really a problem if you are doing an overnight analysis unless somewhere >along the way you're running out of hashspace... I suppose it does happen >sometimes, even for those of us with 192 MB :-). Then the search slows down >to a >crawl. With n=6 the hash tables get full almost six times sooner too or am I >wrong? I suspect Hiarcs already limits the extensions to try to postpone that, >it would be my explanation why you didn't find d5. > >This is probably (just guessing) one of the reasons why Ed hasn't implemented >n-best search yet, to do it really well you'd have to find (n-times) better >algorithms and replacement schemes for not letting the hashtables run over, at >least for n-best overnight analysis. >And especially for customers who have some difficulty saying goodbye to their >old machines that is the only way to go 'really' deep. I'm not saying I myself >wouldn't like such an n-best feature though, but possibly to use it the way I >think you are using it and for which Steve uses Analysis Exclude, to get an >indication which moves are plausible. I would only use it for deep searches >if I could run it comfortably in the back ground for a couple of days. You can't >use a lot of hash, but that would be full quickly anyway. The N-best feature is a very nice one because it produces "real scores". This is certainly a big plus above Rebel's "Analyze Include" and "Analyze Exclude" options. However the disadvantage is that the "N-best feature" is slow as a snail due to alpha/beta. Rebel's "Analyze Include" and "Analyze Exclude" do not have this problem in fact they speed-up the search tremendously. One day the "N-best feature" will be in Rebel too that's for sure. I am still searching for a way to make it faster as I was not satisfied with its speed when I made some experiments. Ed Schroder >I'm not too fond of using Analysis Exclude for deeper searches either, it >means >duplicating a lot of the search each time? I know, only half of the time was >spent on these moves in the last search, but that is still a lot of >information >you would be duplicating then. If you do one Analysis Exclude every night, >when >you are not using the computer for anything else, you get accurate results too >but it still does seem inefficient. > >So what conclusion to draw? I think it is best to always start with a 'normal' >search and for that I'd really like the Overnight Search which is to come with >Rebel Century. Claudio Bollini's diagram of its efficiency was very >impressive. >Thanks for the reviews, Roy, Claudio! >The examples help a lot. I hope we can expect some more? > >After this first search the best way to go forward still has to be invented I >think. >I'd prefer to do a normal search and if there are other moves that look >plausible I'd rather carry them out on the board and start another search, now >one ply deeper. Luckily Rebel doesn't have too much asymmetry in its >evaluations, so after reversing the sign you have a good indication. And if >I'm >not using Rebel ECTool I can make use of Analysis Include with one move to >include, so there should be no asymmetry at all. >It is more risky this way (Analysing not using Analysis Exclude I mean), >but you >have to make a choice which moves to look at somewhere along the way anyhow. >Still this approach is not the best one possible because the computer has >already generated a lot of information about these moves in its first search. >You are not using any of it in this way. There must be better approaches... >Does anybody think some form of temporary position learning could be used >here? >My idea would be to have a separate register for each of what I call the > >non-principal rootmoves, in which to store just the deeper evaluations that >arise after this move. Everything can be thrown away after leaving Analysis >Mode. > >Regarding Rebel, I think it would really be already a useful help if, at >the end >of a search, using Rebel's warroom, the last warroom variations could also be >written to the logfile. And it would be even more wonderful if there were >alpha-beta scores for these moves... > > >Regards,Eelco. > >On August 08, 1999 at 10:33:55, Paulo Soares wrote: > >>On August 07, 1999 at 17:04:31, Eelco de Groot wrote: >> >>>Paolo, I believe you are not the first to ask for this kind of feature :-). >>>Maybe this would mean a major rewrite of Rebel's engine or even a totally >>>separate N-best Engine? As it is written in assembly language, this would >be a >>>tough job. Maybe if you ask Christophe very nicely one day....(ChessTiger is >>>written in C). Actually I have never tried this kind of feature, isn't it >very >>>slow? >>> >>>Regards,Eelco. >>> >>>On August 07, 1999 at 08:49:47, Paulo Soares wrote: >>> >>>>On August 07, 1999 at 04:19:48, Tina Long wrote: >>>> >>>>>On August 06, 1999 at 12:39:15, Paulo Soares wrote: >>>>> >>>>>>Whithout doubt: Hiarcs7.32 >>>>>> >>>>>>Paulo Soares >>>>> >>>>>I agree, >>>>>The Fritz GUI allows: excellent "Correspondence Analysis", printable >gamefile >>>>>with annotations, backwards with English (verbous) game analysis, good >editable >>>>>book, great database features, etc etc. >>>>> >>>>>Hiarcs is currently the smartest engine for the CBase family, Fritz is >much >>>>>faster & deeper. The slowsmart vs fastdeep is roughly equivalent at >Tournament >>>>>time, but (I feel) Hiarcs does better for long time analysis. >>>>> >>>>>Chessbase 7 does all this & a whole lot more, but is extremely expensive >>>>>compared to F5.32 or H7.32 >>>>> >>>>>Meanwhile Rebel Century may be worth waiting for. Ed says it's got a >special >>>>>"overnight" analysis level. >>>>> >>>>>Hi guys, >>>>>Tina Long >>>> >>>>Tina, Rebel is a great program for analysis too, but I feel lack of >>>>the "Infinite Analyses" feature (N moves). I like to analyze positions >>>>looking at some moves, with its variants and evaluations. When Ed >>>>goes to put this feature in Rebel? >>>> >>>>Paulo Soares >> >>Eelco, I never had made a comparison between "normal" mode and "infinite >>analyses" mode. I was curious and I made the following experience: >> >>1. System: PII 300, RAM=192Mb, OS WIN98. >> >>2. Program: Hiarcs7.32, HT=64Mb, Position Learning=OFF, Playng Style=Normal, >> Selectivity=5, Contempy Value=15, Tablebase=OFF. >> >>3. Position: LCTII, Pos14, key move= d5. >> FEN "r2qkb1r/1b3ppp/p3pn2/1p6/1n1P4/1BN2N2/PP2QPPP/R1BR2K1 w kq 0 2", >> >>4. Time to finish depth 9: >> 4.1. Normal: 309"(5min 09sec) >> 4.2. Infinite Analyses(N=6): 1000"(16min 40sec) >> >>5. Move chosen in the end of depth 9: >> 5.1. Normal: d5 (Evaluation= +0.55) >> 5.2. Infinite Analyses(N=6): a4 (Evaluation= +0.41) >> >> >>Obs: I initiated each test after to give a boot in the machine. >> >>In the experience that I made the "Infinite Analyses" mode it's really >>very slow, and the key move, d5, it was not found. >>Independent of the result of that experience, I like to analize >>positions using the "Infinite Analyses" mode, forcing the program >>to make a move, with satisfactory results. >> >>Paulo Soares > > > > > > Re: N-best feature for Rebel - Paulo Soares 00:22:16 8/10/99 (0) > > > [ Post Followup (without quoting) ] [ Post Followup (with >quoting) ] [ Computer Chess Resource Center ] [ CCC Home Page ] >
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.