Author: Robert Hyatt
Date: 08:59:40 07/09/02
Go up one level in this thread
On July 09, 2002 at 03:45:52, Christophe Theron wrote: >On July 08, 2002 at 23:21:06, Robert Hyatt wrote: > >>On July 08, 2002 at 14:11:19, Christophe Theron wrote: >> >>>On July 08, 2002 at 13:27:15, Robert Hyatt wrote: >>> >>>>On July 08, 2002 at 12:48:58, Sune Fischer wrote: >>>> >>>>>On July 08, 2002 at 11:34:36, Robert Hyatt wrote: >>>>> >>>>>>>I too am a DB fan. Just like Bob. >>>>>>> >>>>>>>But I actually agree with you here. I don't think DB did anything >>>>>>>*spectacular*. >>>>>> >>>>>>I totally disagree. Their speed _was_ "spectacular". And that was _the_ >>>>>>point of Deep Blue, after all. Not the point everyone _wants_ to be the >>>>>>point of deep blue, but _the point_ the team developed over 10 years... >>>>>> >>>>> >>>>>Here is a crazy thought, why not simulate DB? >>>>>Given all the papers, I think it should be possible to modify Craft to use the >>>>>same eval and extensions. We turn off hashing, nullmove, SEE and whatever DB >>>>>didn't have. Then we find a slow machine for Tiger and a super fast one for >>>>>Crafty, so Crafty (in DB-mode) has a 200 nps fold advantage. >>>>> >>>>>Ok lot of work, but seems this is the never ending story :) >>>>> >>>>>-S. >>>> >>>> >>>>This would be great if we had some of the DB guys helping. Unfortunately, >>>>while they revealed a lot about various parts of DB, there is no single >>>>comprehensive source paper to use as a reference. IE what are those 8,000 >>>>unique eval terms in DB (some of those terms actually represent a matrix with >>>>multiple values so it is actually more complex than that)? >>> >>> >>> >>>Sorry but the "8000" includes every entry of every matrix. >> >>Not according to the things I have seen written. But it really doesn't matter >>to me either way. I don't have anywhere _near_ 8000 terms in my evaluation. >>I don't have 1000 unique terms, even counting all the piece/square tables. >> >> >> >>> >>>It's like saying that a piece square table program is composed of 768 unique >>>eval terms (64 squares x 6 piece types x 2 colors). >> >>Even if that were done, that is only 10%. What about the other 90%? You >>have a _lot_ of counting to go to reach 8000... > > >They say in their paper that many terms were not used. > > > > Christophe > Correct. They also said that 8000 _were_ used. Hsu has said that maybe 50% of the total evaluation hardware was actually "turned on" in 1997 due to time constraints. > > > >> >> >>> >>>If I count this way, I guess that Chess Tiger must have something like 50000 >>>unique eval terms... :-) >>> >>> >>> >>> Christophe >>> >>> >>> >>> >>>> Ditto for some of >>>>their search algorithms. They have given lots of 'hints' about things, but >>>>significant implementation details are not available. >>>> >>>>IE something like trying to build a F-1 by looking at it run around the track. >>>>There are _significant_ details that are not readily apparent from such >>>>observations...
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.