Author: Jesper Antonsson
Date: 15:10:20 10/21/02
Go up one level in this thread
On October 19, 2002 at 18:16:17, Uri Blass wrote: >On October 19, 2002 at 17:59:05, Jesper Antonsson wrote: >>Why? If they reached 12 true plies (without nullmove) and did more extensions on >>top of that, why do you "strongly believe" that that was not "more" than Fritz's >>14 pruned plies? > >I believe based on analysis of the games. > >I found that Fritz on p800 was only slightly slower in seeing the same line >relative to deeper blue in few cases. All the time? There are few games and perhaps the positions were such that DBs superior positional and tactical strength didn't show that much. >Hsu also did not claim that they did 126M effective nodes. > >He said that the estimate for the real number of nodes(not efective nodes is >200M and with all their problem of parralel search and hardware it is logical to >assume that the eqvivalent number of nodes on a single machine is a lot more >than 2 times lower. Half of 200M is 100M. Still more than SMP micros. And I don't think he threw his NPS at nothing. These are smart guys. > And remember that DBs eval was much better than Fritz's. (I >>know you won't agree, but I believe Hsu here, because I know the DB team was >>smart and because Hsu in a better position to know than you are.) > >Hsu knows nothing about the evaluation of Fritz How do you know? From the interview, it seemed he has followed the DF-Kramnik match, and he mentioned the positional superiority of a dumbed down DB Junior that recognised worthless open rook files and so on. Does Fritz handle the things he mentioned well? >and it make sense to believe >that Fritz's evaluation is better because Fritz had more time to test their >evaluation. That's just one factor. When some PhDs collaborate with GMs and devote a year to fine tuning the eval of DB, using some clever tools they devised earlier, such as automatic tuning, it can become very good. Also, something that has been reiterated again and again, they had the possibility to do things in hardware for free, things that are too costly for software implementation. They could implement such very complex eval terms in hardware, and if they were properly parameterized, they could then spend considerable time tuning it. I obviously can't be sure that it was(is) superior, but I'm firmly convinced that it could have been made superior, and think Hsu et al were smart enough to do it. /Jesper
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.