Author: Uri Blass
Date: 11:32:14 08/21/02
Go up one level in this thread
On August 21, 2002 at 13:50:18, Robert Hyatt wrote: >On August 21, 2002 at 07:57:50, Uri Blass wrote: > >>On August 21, 2002 at 06:33:05, Gordon Rattray wrote: >> >>>On August 20, 2002 at 21:48:59, Robert Hyatt wrote: >>> >>>>On August 20, 2002 at 18:42:07, Uri Blass wrote: >>>> >>>>>On August 20, 2002 at 17:49:34, Robert Hyatt wrote: >>>>> >>>>>>On August 20, 2002 at 17:02:52, Uri Blass wrote: >>>>>> >>>>>>>I include only gam,es from 1993 >>>>>>> >>>>>>>Deep Blue - Hamann,S >>>>>>>5R2/r5pp/3N2k1/3Ppp2/2P5/p7/P5PP/2n3K1 w - - 0 38 bm Nb5 >>>>>> >>>>>>What did DT play here? I only find Nb5 and I find it instantly and stick >>>>>>with it... score =+1.5 so apparently DT was winning here. Did it lose by >>>>>>playing something else??? >>>>> >>>>>Yes it lost by Rxf5 >>>>> >>>>>This is a typical computer move because it wins a pawn if you do not search deep >>>>>enough when Nb5 does not win a pawn(unless you search very deep and I am sure >>>>>that your +1.5 is based on poasitional factors). >>>> >>>>The problem here is that we don't know what happened in the game. IE deep >>>>thought _always_ played over a network connection. In the late 80's and >>>>early 90's the net was not as accessible and reliable as it is today, and it >>>>is certainly possible that they lost the connection, didn't know it, and >>>>after they found out, they had to move quickly. Happened to me and Cray Blitz >>>>many times. Happened to deep thought against Fritz in Hong Kong 1995 in >>>>fact... >>> >>> >>>I think I remember reading about this in a chess magazine (Chess Monthly, UK) at >>>the time. DT played Rxf5 after about 30 seconds. Later it was discovered that >>>a 42 second search would have found Nb5. Even if my figures aren't exact, I'm >>>pretty sure they're close (I can dig it out if necessary). I don't know what >>>the time control was for the game. >>> >>>Gordon >>> >>> >>>[snip] >> >>I read about it in the past and if I remember correctly Deep thought needed 48 >>seconds to find Nb5 but in the game(1 hour per game) it used only 42 seconds. >> >>The point is that most of the top programs of today can see it faster(some >>programs like crafty even see it immediately). > >The question is, what is the program seeing? IE Crafty might play it due >to positional considerations that might well be right or wrong. DT might >not have played it for the same sort of positional reason. IE it could have >played the move it played due to others looking worse, or to that one looking >better, we don't know... > >I can show you a couple of horrible moves my program played as well. And I >can tell you _why_ it played them. In 1978 we played a horrible move against >Belle at the ACM event. Because unknown to us, our Univac machine (on loan to >me for the event) had a memory problem and had taken 1/4 of memory offline. >That caused me to get swapped out more than once when a system task had to run, >and we made one move after taking 2+ minutes of clock time, but 0.01 seconds of >cpu time... In 1980 we dropped a pawn against Chaos on a Cray. Because some- >one ran something on our Cray that was supposed to be dedicated, and they ran >at high priority and swapped us out for over 5 minutes. We got swapped back >in and made a move without searching a single node. > >Until you know _why_ something was done, it is impossible to say whether that >was an error caused by outside circumstances or programming problems. > > > >> >>I looked only in few games from 1993 and I do not expect to see the same number >>of mistakes with top programs of today. >> >>If you look in many games you can find also mistakes of tiger that other >>programs can avoid but I believe that it is possible to prove that the number of >>tactical mistakes that deep thought did in games is simply bigger. > >I don't think so. Because of the analysis I did on the 1986 Cray Blitz >games using Crafty of today. It didn't find a _single_ move it would label >"tactical mistake". Not a one. They didn't always agree on the final move >chosen, of course, but they never disagreed by more than a pawn in real >material... If we weren't making horrible tactical errors in 1986, then >DT certainly wasn't doing so at a speed that was 10X what ours was... If your opponent are 400 elo weaker relative to you then you can often beat them without tactical mistakes. The interesting games are games when the difference is smaller than 200 elo. I chose humans with the relevant level. It is possible to do the same experiment and use games of Junior,Fritz,Shredder,Tiger against humans(Smirin, Gulko). Uri > > > >> >>I did not have the time to check it so I may be wrong but the only way to >>convince me is to show data and my impression based on looking at games is that >>deep thought blundered more often. >> >>Uri
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.