Author: Robert Hyatt
Date: 15:14:00 07/07/99
Go up one level in this thread
On July 07, 1999 at 14:38:34, blass uri wrote: >I looked at my database for Cray blitz games and I was surprised by the weak >results of cray blitz. > >I expected better result after I read the results of 100% against Genius1. > >I have 5 games of cray blitz from 1994 and it lost 2 of them against Wchess and >Zarkov and drew against Now(It won 2 games against innovation and spector). > in the 1994 tournament, we had a really bad bug. That was the first version where we ran with _full_ singular extensions implemented. I had a pure FORTRAN version that I tested with, and then assembly code that we used to play games with. The assembly code was the same as the FORTRAN in most places, but in one key respect it wasn't... The FORTRAN code had a max ply limit of 64, which was the deepest we could search without blowing out arrays. Older versions of the program never reached this limit except with a bug, so we left the maxply check in the FORTRAN, but Harry removed it from the 'production' assembly code as it slowed it down a percent or so. In 1994 we were blowing out the top of those arrays badly until we finally went back to an old version in the last game. In the WchessX game, we had queens magically appearing on the board, program crashes, etc. Always in positions where there were long 'forcing' tactical lines that the singular extensions followed very carefully. That was why I dropped the "Cray Blitz" project... we never had enough machine time to test to catch such problems... At the ACM event, by the time we had arrived, we had played _no_ games with the new code... just a few test positions, mostly on much slower hardware. Where we couldn't get to depth=60 PVs at all.. >I have 5 games from 1993 and cray blitz won 2 games(against Now and Mchess) and >drew 3 of them (against Startech,Gideon and Socrates). > >I have 4 games from 1991(I do not count the game against Deep thought) and >Crayblitz drew 2 of them against Socrates and Gideon and won 2 against Zarkov >and BP. > >I do not see something close to 100% result and I do not think that programs >like Wchess were better than Genius and even if I assume that they are slightly >better then 9 out of 14 is a disappointing result. > >Uri The results were bad. From 1980 to 1986 or so, we got plenty of time to run lots of tests. After the late 80's, we were lucky to get machine time to play in the real ACM event. You can _not_ do well with a program that is tested like that. Once we fixed the singular extension bug in 1995 (I think) it played like its old self again, but who knows what other kinds of bugs would crop up with as little testing as we did. As I said, winning any WCCC or ACM event takes a certain amount of luck. We won in 1983 without ever having played a game with our brand new two-week-old parallel search...
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.