Author: Robert Hyatt
Date: 08:04:35 09/18/05
Go up one level in this thread
On September 18, 2005 at 07:53:16, Rolf Tueschen wrote: >On September 18, 2005 at 01:44:33, Uri Blass wrote: > >>On September 18, 2005 at 00:51:17, Robert Hyatt wrote: >> >>>On September 18, 2005 at 00:36:15, Mridul Muralidharan wrote: >>> >>>>On September 17, 2005 at 23:53:05, Robert Hyatt wrote: >>>> >>>>>On September 17, 2005 at 22:33:58, Vincent Diepeveen wrote: >>>>> >>>>>>On September 17, 2005 at 17:02:09, ALI MIRAFZALI wrote: >>>>>> >>>>>>Take a time control of 40 in 2 with pocket fritz against crafty without nullmove >>>>>>and give crafty factor 1000 in nps extra. >>>>>> >>>>>>Deep Blue didn't use nullmove either. >>>>> >>>>>What does null move have to do with it? >>>>> >>>>>In 1996 there wasn't a computer on the planet that could beat deep blue. This >>>>>is almost 10 years later. >>>>> >>>>>What is the point of this discussion??? >>>>> >>>>>Just a very lame attempt at starting a flame war, about a statement I supposedly >>>>>made? (a statement I did _not_ make by the way)... >>>>> >>>> >>>>I dont know about the flamewar part (I thought Vincent did not start the thread) >>>>, but the 'without null move part' might be referring to the fact of deep blue >>>>not using it. >>>>Actually a good test would be : >>>> >>>>1) 1000x nps advantage >>>>2) No null move >>>>3) Use full singular extension as 'explained' by them. >>> >>>This is a completely worthless experiment. Take out my null-move search. >>>Attempt to graft their singular extensions onto my program. What about my >>>evaluation? My search extensions? How could one possibly add and remove bits >>>and pieces of Crafty, to make something into the approximate skill of deep blue? >>> >>>Next, why is this important? My 1000x statement had nothing to do with >>>null-move vs no null-move... >>> >>>> >>>>I suspect (3) _will_ kill your search and keep the searchdepths much below what >>>>the pocket fritz will get :) >>> >>>I don't think so. I had their full SE implemented in Cray Blitz. Its cost was >>>almost 2 plies. But then tactically it was reaching very deep stuff to offset >>>that. I've never found a workable SE that impressed me as "this is really good" >>>when it comes to Crafty... >>> >>>On the quad opteron, my search depth would then probably drop to 13 plies. But >>>then I get a factor of 1000X faster. My branching factor would be closer to 6 >>>with no null-move, which would ramp me up by 4 plies without null move, or 10 >>>plies with normal null move and a branching factor of around 2.0... >>> >>> >>> >>>> >>>>So , even though Deep blue might have been invicible from programs of that age's >>>>standards - it will get royally kicked around by even weak amateur programs of >>>>today (bugfree ones ofcourse) running on modern hardware ! >>>> >>>>Regards, >>>>Mridul >>> >>>again, based on what? Null-move is not generally credited with making a program >>>200 Elo stronger. I might one day run some decent-length games with null vs >>>non-null to see what the actual rating difference would be for Crafty, then one >>>might actually extrapolate what 1000x faster hardware would do by actually >>>playing that time-odds match as well. Then we don't need to guess, speculate, >>>or anything else... >>> >>>I'll remind you that a few years ago (I don't remember exactly when although you >>>can find precise mention of the experiment here) I had the chance to play Crafty >>>(I think on some quad box, which one I really don't remember) against Cray Blitz >>>(less singular extensions, the version on the machine I had access to did not >>>have that version) on a T932. Thing was searching about 7M nodes per second. >>>It gave crafty a pretty good drubbing. And Cray Blitz on that machine could not >>>touch deep thought, much less deep blue, based on actual OTB games against them >>>at ACM events... >> >>I think that saying that it could not touch deep thought on that machine is >>misleading because I remember that Cray blitz that played on tournaments >>searched significantly less than 7M nodes per seconds. >> >>If you want to claim that latest Cray blitz on the machine that searched 7M >>nodes per second was probably weaker than Deep Thought then you cannot use data >>about tournaments when Cray blitz searched 200K nodes per second or 500 Knodes >>per seconds. >> >>Uri > > >Bob, >you should give us a couple of statements about possibilities to compare >something with chess programs and machines. > >Also, your sentence above is really difficult to say the least: > >"against Cray Blitz (less singular extensions, the version on the machine I had >access to did not have that version) on a T932." > >Could you clarify what is "version" and "version"? - BTW I've understood it now >after the third read. Your test version hadn't this with less singular >extensions, what Crray however had formerly. Is that correct? In either 1993 prior to the ACM event, or else right after the ACM event, I added singular extensions as laid out in the Hsu/Campbell paper. We definitely played in 1994 with them, and I am almost certain we played with them in 1993, based on a bug I believe happened in 1993 (we extended so deeply in a passed pawn ending that we overflowed an array and blew up in one key position. we had to re-start and managed to play on and continue the game..) The T932 test I ran was on a machine being "phased out" (Can't say where). Friend of mine asked "we are turning this thing off in a few days, but no one is using it. Want to test to get some SMP data with 32 cpus?" This was after the point where I had lost all the source files, but he had an older version on the machine already, and I played around with that. Hence the "non-singular-extension version comment". The version I used was 3-4 years older than the most recent SE version of Cray Blitz. The reason then, was that I could either play with the version he happened to have in the user's library files, or not at all, since I had no source files of any kind for Cray Blitz, except for a big print-out and I was not going to re-type 50K of FORTRAN and 20K+ of cray assembly code... So I think your interpretation is correct. It was a poorly worded sentence on my part... Hopefully the above makes it clearer... > >Further I want to support Uri with his questioning of comparing something actual >with something weaker from ACM tournaments. Or did you mean that you re-played >and analysed your former games with the actual version (of Cray Blitz)? First, in my response to Uri, I pointed out that we used a C90 in 1993/1994, and we also used a T932 for a couple of games, when the machine was available. Usually it happened in the early rounds of the tournament, which were played on a Sunday, vs the later rounds (ACM/WCCCs were 5 rounds back then) that were played on week nights during "prime time". Second, I am not certain I could accurately recall what machine we used on what round. The NPS quoted in the ACM tournament press releases was usually wrong, because we had to enter 6 months prior to the event, and we often didn't know which machine or machines we would be able to use until the week of the tournament. The fastest we ever searched at an ACM event was about 3M nodes per second on an 8-cpu T90. Rough speeds on Crays were 200-300K on a YMP, 500K-1M on a C90, and up to 7M on a T932, which I think we only got to use in one game. A T90 we used developed a problem on one cpu dealing with negative integer math, and would produce wrong answers. Got worse and worse until we had to get off that and move back to a YMP that could at least produce correct results. That is why I said it was hard to specifically give NPS numbers for Cray Blitz. In one event we played X rounds on one machine, Y rounds on another, and one round on yet another, due to scheduling issues. That is one reason I chose to depart the supercomputer platform after 1994, it was just too big a hassle compared to what I see today... :)
This page took 0.01 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.