Author: Robert Hyatt
Date: 14:16:04 03/03/00
Go up one level in this thread
On March 03, 2000 at 15:18:09, Dave Gomboc wrote: >On March 03, 2000 at 08:09:47, Robert Hyatt wrote: > >>On March 03, 2000 at 03:44:48, Dave Gomboc wrote: >> >>>On March 02, 2000 at 18:26:09, Robert Hyatt wrote: >>> >>>>On March 02, 2000 at 01:42:43, Ed Schröder wrote: >>>> >>>>>On March 01, 2000 at 20:52:52, Jonathan Lee wrote: >>>>> >>>>>>Thanks, Pete on IBM's logic on their selling of chess software. If their >>>>>>software was awesome, they already would have a large share against the top >>>>>>guns. >>>>>> >>>>>>Someday said every 18 months the hardware doubles in speed. >>>>>>Someone else said IBM's hardware "DB" is 1000 times faster than the 500MHZ PIII. >>>>>>In the year 2015, (what a bold prediction) PC's will be as fast as Deep Blue. >>>>>>Do the math: >>>>>>2 ^ 10 = 1024 (twice the speed 10 times equals about 1000 times faster) >>>>>> >>>>>>18 * 10 = 180 (18 months multiplied by ten times equals 180 months) >>>>>> >>>>>>180 months equals 15 years >>>>>> >>>>>>2000 + 15 = 2015 (We are now in the year 2000; add 15 years and you will get >>>>>>2015) >>>>> >>>>>What about today? >>>>> >>>>>Nowadays programs hit a deeper ply depth than DB. See the log files >>>>>on the IBM site. >>>>> >>>>>Ed >>>> >>>> >>>>No... you are mis-reading the logs. When you see depth=10/5, that means >>>>10 ply (full width + extensions) in the software, another 5 plies + everything >>>>but singular extensions in the chess engine. When you look at those numbers, >>>>we are getting crushed in search depth. 10(5) means 15 plies full-width, no >>>>null-move or anything, + their extensions. >>>> >>>>This was covered when we were discussing the DB logs. I confirmed this >>>>point with two different DB team members to be sure we were understanding the >>>>(n) number correctly. Looking at the logs, the hardware searches 4-6 plies >>>>depending on the base software search depth. When you see 11(6) that is a >>>>full-width 17 ply search, which is awesome. >>> >>>Double-check that. 11(6): 17-ply search, yes. _full-width_ 17-ply search, I >>>don't think so. The top of the software search is full-width, the bottom of the >>>software search is selective, and the hardware search is full-width again. At >>>least, that is my understanding. The interesting question is how many of the 11 >>>in software were full-width... and my recollection is that the answer is not 11. >>> >>>Dave >> >> >>There is _no_ selective search in DB. They do "selective extensions" in the >>software, but _every_ move is searched to at least the depth given. IE if you >>see 11(6), then *every* move at the root is searched to at _least_ 17 plies >>deep, not counting the capture search. And most are searched deeper due to the >>extensions. >> >>_None_ are searched to less than that depth. DB _never_ had any sort of forward >>pruning in the normal search, although it did have a sort of futility cutoff in >>the capture search only. >> >>This is all well-documented. Hsu didn't believe in selectivity at all when >>applied to pruning. >> >>The answer to your question _is_ "11". > >It's well-documented for DT. Where is it well-documented for DB? I know Hsu >didn't like pruning, but by itself that doesn't mean that they didn't do any in >DB. Does he say so in his book? > >My recollection of a conversation with a member of the DB team in 1997 is that >there was some selectivity going on in the software search, in the manner that I >described above. > >Dave All I can say is what I have been told, which is that "the only selectivity we do is in the extensions, _not_ in forward pruning." In his book (or in something he posted somewhere, maybe the IEEE article) he mentioned that he would like to try something selective (like null-move) in his _next_ chip, when he was talking about doing one. That suggests to me that he isn't in the current DB...
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.