Author: Robert Hyatt
Date: 07:08:15 07/22/00
Go up one level in this thread
On July 22, 2000 at 05:44:36, Ed Schröder wrote: >On July 21, 2000 at 22:27:45, Robert Hyatt wrote: > >>On July 21, 2000 at 19:16:41, Ed Schröder wrote: >> >>>On July 21, 2000 at 15:29:26, Robert Hyatt wrote: >>> >>>If you don't mind I only answer those points not earlier discussed >>>(enough) to avoid ending up in endless circles. >>> >>> >>>>>2) DB is no brute force program (as you always have claimed). Quote >>>>>from the IBM site: >>>>> >>>>> "Instead of attempting to conduct an exhaustive "brute force" >>>>> search into every possible position, Deep Blue selectively >>>>> chooses distinct paths to follow, eliminating irrelevant searches >>>>> in the process." >>>>> >>>>>I always said this after I had seen the log-files. It beats me how you >>>>>always have claimed the opposite on such a crucial matter presenting >>>>>yourself as the spokesman of Hsu even saying things on behalf of Hsu >>>>>and now being wrong on this crucial matter? >>>> >>>>Sorry, but you are wrong and are interpreting that wrong. DB uses _no_ >>>>forward pruning of any kind, this _direct_ from the DB team. The above is >>>>referring to their search _extensions_ that probe many lines way more deeply >>>>than others. If you want to call extensions a form of selective search, that >>>>is ok. It doesn't meet the definition used in AI literature of course, where >>>>it means taking a list of moves and discarding some without searching them at >>>>all. >>> >>>The quoted text describes DB as a selective program, no brute force. I >>>don't see how you can explain it otherwise. The text is crystal clear. >>> >>> >> >>Why don't you simplyh ask Hsu, or are you afraid you will get an answer >>you don't want? DB was _always_ brute force. Every document written about >>DB said this. The paragraph you are quoting is talking about "selective >>search extensions" which was one of the real innovations from the Deep Thought >>development (singular extensions, later used by Lang, Kittinger, Moreland, >>Hyatt, who knows who else). > >I disagree. Extensions are always selective. Some moves are extended >some don't and that makes that extensions is a selective process by nature. >So the text (about brute force) can't be related to the previous sentence >(about extensions). They made 2 statements (not one). > There is a difference. It is one thing to search move A one ply deeper than move B, based on some (hopefully good) criteria. It is quite difference to simply choose to take move B and not search it at all from the current position. One easy example is that in the q-search, I throw out _all_ non-captures and consider them no further, while captures continue to grow trees below them... > >>You _know_ they were basically in the same mold as the rest of us. This has >>_never_ been in doubt. >> >>If you do doubt it, just ask the horse's mouth, since you don't want to believe >>me. >> >> >> >>> >>>>This _was_ deep thought. It was doing about 2M nodes per second in 1995, >>>>according to Hsu. >>> >>>Then Hsu is wrong or the IBM site. >>> >>>Quote from the IBM site: >>> >>> "Deep Thought acquires 18 >>> additional customized chess >>> processors and emerges as >>> Deep Thought II. It now is >>> running on an IBM/6000 and >>> can search six to seven million >>> chess positions per second. >> >> >>That was correct. But as I said (after a conversation with Hsu) it _never_ >>really ran at that speed. The few times they tried to use all the hardware, >>things didn't work out very well (this was mainly used during the Fredkin >>stage II matches, where they physically shipped the machine (a single Sun >>workstation + the VME cards) to remote locations. >> >>Hsu has said point blank, the most recent version of DT was searching about >>2M nodes per second. I take him at his word, since he built the thing... > >The only thing that counts here is the contradicting data: > >1991: IBM 7 million >1995: Hsu 2 million > >Now who to believe that's the question. > > Simply ask Hsu. wouldn't you think??? I used the 7M speed in a post once (either here or in r.g.c.c) and he sent me a private email correcting the number. I believe 7M was the peak number, while 2M was the _effective_ number and matched the 200M number from DB. As I said before, everybody used to report MAX or TYPICAL NPS, but the number was RAW. IE for Cray Blitz, my 8 cpu numbers were 8x my one cpu numbers. Hsu changed the way he reported this so that the numbers were more realistic. IE with CB you might conclude that it would search to the same depth in 1/8th the time, since the RAW NPS was 8X the one CPU number. That didn't happen. With Hsu's numbers, if a single chip went 2M, he says that 480 CHIP DB won't search the same tree in 1/480th of the time. Rather it will search it in 1/100th the time (2M / 200M, rather than 2M / 1B. If you want to use 7M for DT, then lets use 1B for DB2, as that is comparable. If you want to use 200M for DB, then 2M for DT is the right number. All right from the "horse's mouth" if you know what I mean... I don't believe he started using the effective number until his PhD dissertation, which was well after DT.02 was out and running... > >>> >>>6 to 7 million NPS. This in the year 1991 so 4 years before the Hong Kong >>>event. So according to Hsu and/or IBM in 1995 the machine dropped from 7 to >>>2 million NPS?? One might expect the opposite, a faster machine after >>>4 years but not a slower one. Something ain't right with these numbers. >> >> >>Simply email Hsu... it was his box. He can tell you what you want to >>know... >> >> >>> >>> >>>>Fine. Again, Hsu is a liar. If that is what you want to think. Here is >>>>an excerpt from him that might help: >>>> >>>>=============================================================================== >>>>Web-based DB Jr uses a single card, a random opening book (including >>>>fairly bad lines) and one second per move (a quarter of which is used >>>>in downloading the evaluation function, and the search extensions are >>>>more or less off due to the very short time). It probably plays at around >>>>2200, which is usually sufficient to play against players in random marketing >>>>events. Repetition detection is also turned off (The web-based program >>>>is stateless). The playing strength of "DB Jr." spans a quite wide range, >>>>depending on the setup. The top level, which we used for analysis and >>>>in-house training against Grandmasters, is likely in the top 10 of the >>>>world. >>>>================================================================================ >>> >>>I said the contradiction is in the private emails so you can't know. >>> >>>Ed >> >> >>No, but I believe from the above, which is also private email, there is >>absolutely no confusion in what "web DB Jr" was. It is _very_ clear, and >>not open to misinterpretation, wouldn't you say?? >> >>It was thrown together at the request of marketing guys. And "thrown together" >>is a pretty accurate description. He says "2200". In another email he said >>"2200 might have been optimistic"... > >Every time it is something else. I stopped believing it. > >Ed That is your choice of course. I know him a bit better...
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.