Author: Robert Hyatt
Date: 06:53:14 09/16/04
Go up one level in this thread
On September 16, 2004 at 09:20:05, martin fierz wrote: >On September 16, 2004 at 08:39:13, Robert Hyatt wrote: > >>On September 16, 2004 at 08:19:20, martin fierz wrote: >> >>>On September 15, 2004 at 12:18:42, Robert Hyatt wrote: >>> >>>>On September 15, 2004 at 10:32:53, martin fierz wrote: >>>> >>>>>On September 15, 2004 at 09:53:53, Stuart Cracraft wrote: >>>>> >>>>>>Hi, >>>>>> >>>>>>Anyone know of some code somewhere that implements >>>>>>at least part (or all) of the originally described >>>>>>singular extension and/or any modifications to it that >>>>>>have proven worthwhile (if any)? >>>>>> >>>>>>I am curious what mediocre (or better) results people >>>>>>have gotten with singular extension. Originally Anantharaman >>>>>>hypothesized that it wouldn't be good at the slower >>>>>>speeds of most programs at the time and would require >>>>>>fast speeds to show effect. Has this proven true or >>>>>>false in the intervening 15 years? >>>>>> >>>>>>Is singular extension now generally discredited as a >>>>>>non-reproducible singularity in and of itself? >>>>>> >>>>>>Thanks, >>>>>> >>>>>>Stuart >>>>> >>>>>AFAIK, SE is 'interesting' in the sense that it does enable programs to solve >>>>>certain positions faster, but of course you pay a price. and again AFAIK, nobody >>>>>is really using it these days, because the price seems too high to pay. i.e. in >>>>>games it's no improvement. >>>>> >>>>>just because the deep blue team used SE doesn't mean it's any good. remember, >>>>>they also decided not to use null-move, which was an established concept by >>>>>then. >>>>> >>>>>cheers >>>>> martin >>>> >>>>Remember also that _others_ use/used SE. Cray Blitz did starting in 1993. >>>>Wchess (Kittinger used the PV-singular half of SE.) I suspect others did/do as >>>>well. IE we know that Ferret had an implementation of SE. >>> >>>of course others used SE. if deep blue had been using *anything*, others would >>>have tried it too. like everybody playing the kings indian after kasparov did... >>> >>>i have never seen anybody claim that SE is of any use. >> >> >>1. It worked for CB. > >how do you know it worked? Lots of testing, and two ACM tournaments actually running on the big iron... I ran thousands of games on slower hardware while working on DTS... > >>2. Kittinger used it for several years in Wchess and claimed it worked well for >>him, using a "partial implementation". > >how do you know it worked? I saw it play. I saw its PV displays as Dave would automatically turn it on when playing Crafty as I did when playing him. And we compared notes on some sticky test positions. And finally it was in the _distributed_ program version, which would suggest that he thought it was good. :) > >>3. Ferret used it. > >how do you know it worked? > Bruce has reported _lots_ of test data here in CCC. Including ECM results with and without, etc... >>4. HiTech used it. > >how do you know it worked? Are we stuck in a loop here? They produced good results in games against computers and humans. I'd trust their test methodology. > >ok, ok, i'll skip the rest :-) >these are all private programs, where nobody can verify independently what it >helps or what not. SE may be fun to play around with, but has anybody ever stood >on a soap box and said "my program gains XXX elo by SE"? Yes. Ferret. HiTech. Deep Thought. Deep Blue. Don't know if Dave claimed any particular Elo improvement, but he did claim that "PV-singular was an improvement tactically and made his program stronger." Cray Blitz was better, as verified by testing by myself. I never quantified it in terms of raw Elo, and it was not a _huge_ improvement. But it was better. >this statement is different from "it works". i have stuff in my program that >"works", but i can't put down a number as to whether it really helps or not. >generating checks in qsearch is such an example. my program finds many things >much earlier in ply-terms and earlier in time with checks in QS. does it help >actual play? i don't know... but it works in the sense i wanted it to work. > > >>That conclusion is flawed. It says more about the _implementation_ than it does >>the idea itself. Others have had success with it as I mentioned above. > >others have said they had success with it. that is not the same. i have seen no >proof anywhere, and as MPC-crafty suggests, you should never trust the guys who >invented an algorithm. i know of no independent tests. > >cheers Bruce tested for weeks and weeks, reporting games, test suite results, etc. Commercial guys don't reveal such results of course... I tested for a year on our old 30-processor Sequent Balance 21000 when working on this particular idea.. By comparison I don't think you get 30+ elo from any single extension. For example I doubt the recapture is worth that. But testing Crafty with and without showed it to be better with... Ditto for Cray Blitz. YMMV of course as Ken Thompson was the first to use it that I know of, and he later decided it was not very good for Belle. But then Belle was not pure alpha/beta since part of the search was hardware and part software... > martin
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.