Author: Jeremiah Penery
Date: 11:05:48 10/20/02
Go up one level in this thread
On October 20, 2002 at 12:12:26, Uri Blass wrote: >On October 20, 2002 at 11:29:40, Jeremiah Penery wrote: > >>On October 20, 2002 at 11:20:55, Uri Blass wrote: >> >>>On October 20, 2002 at 11:00:42, Jeremiah Penery wrote: >>> >>>>On October 20, 2002 at 03:41:56, Uri Blass wrote: >>>> >>>>>On October 19, 2002 at 21:44:49, Fernando Villegas wrote: >>>>> >>>>>>I have just finished the book by Feng-Hsiung Hsu. In just a shot, from after >>>>>>lunch to this time, meal time. Very interesting. You cannot stop he reading. >>>>>>First big impression: if this guy and his team had worked just one year more on >>>>>>Deep Blue, Garry has been crushed to ashes, to atoms. Yes, because once and >>>>>>again Deep Blue appears as an uncomplete device full of bugs and problems, >>>>> >>>>> >>>>>I am surprised that after this people still believe that it's evaluation was >>>>>better than the evaluation of Deep Fritz of today. >>>> >>>>I'm surprised people still think it's evaluation was worse than Fritz of today. >>>>But it doesn't matter - I think DB will never play again (stupid IBM), so >>>>neither side can EVER win the argument. >>> >>>The main point is the following: >>>"Deep Blue appears as an uncomplete device full of bugs and problems" >> >>They did have a lot of bugs over the course of the development, but what does >>that have to do with the quality of the evaluation? >> >>>I do not believe that something that is full of bugs and problems can have >>>better evaluation than Deep Fritz that was tested more seriously. >> >>Almost all of the 'bugs' were not in the evaluation. > >I understood from the discussion that they discovered bugs including bugs in the >evaluation in games 1 and 2 of the match against kasparov and fixed them for the >rest of the match. There was an evaluation feature on the hardware that was disabled in the first game, which caused DB to think the queen trade was good. I'm not sure it affected the second game, but it's possible. >Maybe I undertsood wrong and I did not read Hsu's book. >I do not want to support Hsu after he left computer chess. > >I think to buy it only if Hsu returns to computer chess. > >> >>Should I say Fritz has big evaluation 'bugs' because it completely misevaluates >>many positions? Should that be cause to say Fritz has worse evaluation than >>some other program(s)? > >No >It is different. > >Fritz may evaluate wrong something because of lack of knowledge but >I believe that the demage from wrong knowledge is often bigger than the demage >from lack of knowledge. Slightly untuned knowledge is not 'wrong knowledge'. For example, if you tell the program that advancing pawns is good in the endgame, it will play better in most cases where that comes up, even if it is sometimes wrong due to poor tuning. Remember, DB evaluation was not completely untuned or untested. Joel Benjamin played a lot of testing games - when he found a position the machine played badly, they fixed it. They also had the automatic evaluation tuner, which tuned the evaluation based on short searches to more closely match GM analysis. Would you rather have a program with no king safety (no knowledge), or a program that over-evaluates it (untuned knowledge)?
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.