Author: Robert Hyatt
Date: 06:57:32 04/24/00
Go up one level in this thread
On April 24, 2000 at 02:10:56, Jeremiah Penery wrote: >On April 24, 2000 at 00:42:47, Christophe Theron wrote: > >>I think we can say that a given amount of knowledge (without taking search into >>account) gives you X elo points. Then the search (actually the depth of the >>search) gives you Y additional elo points. >> >>For a given program, X is fixed and does not depend of the time control. Y >>depends on the speed of the computer and the search time. > >I know of at least one specific case where X is not fixed relative to the time >control. The knowledge is often tied in with the search - i.e., You may have >some evaluation parameters that only affect low-depth searches, and some that >don't actually do anything until you reach much higher depths. The one specific >example is from the testing of Cray Blitz. He added some knowledge that helped >the program while playing at very short time controls (Or, a much slower >machine.), but totally ruined the program when it searched deeper. > >I don't think you can say X is necessarily fixed for every program, because >there are always things you may put in to fix problems you see on your machine, >but if you run on much slower/faster machines, that 'fix' may really break more >than it fixes. Because it's all dependent on the dynamics of the search. I had forgotten about that. The problem dealt with pawn holes. We played many games vs the old supercon machine (Kittinger) and saw (we were on a vax) too many endgames where the opponent could penetrate due to a ragged pawn structure. I added code to stop that, because the program didn't understand holes unless the search could see a way to exploit them (ie a knight in an outpost square, etc.) We stopped losing those endgames on the vax. But when we played in the 1985 ACM event and even the second round of the 1986 WCCC, we were searching about 2x deeper (10 plies rather than 5) and now it became paranoid about holes and would never push a pawn, period. Removing those 2-4 lines of code made all the difference on the faster cray hardware. Adding them made all the difference on the slower vax. I learned to not tune the eval on slow hardware, unless I was reallyu going to play on that hardware.
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.