Author: Amir Ban
Date: 13:20:29 01/02/98
Go up one level in this thread
On January 02, 1998 at 14:59:02, Don Dailey wrote: >On January 02, 1998 at 14:03:48, Stuart Cracraft wrote: > >>Perhaps not directly related but I'll share this anyway... >>My experience in adding knowledge to the evaluation function >>and its relation to test results is that pure tests like >>Win-at-Chess and other tactically-based tests, including >>rating tests that are more tactical (like Kaufman's) suffer for >>more evaluation knowledge but that non-tactical tests may benefit. >> >>A recent result was that correcting some pawn structure logic >>in a program that was mis-evaluating passed/doubled/isolated >>pawns dropped the Win-at-Chess score by about 2.66% in total >>problems solved but raised the Louguet rating by 45 ELO points. >>At the same time the Louguet result went up 45 points the >>Kaufman result went down 24 points. >> You were almost certainly looking at random noise. In a 300 position test-suite, I can get the SAME program scoring 3-6 less on its "bad" day. Test suites like Louguet are so small that no real significance should be attributed to any particular result. Understanding the effects of noise is the key to keeping your sanity in this business. Controlling it is a different question. >>This result came by making only these changes: >> 1. correctly evaluate passed pawns based on rank >> (previously, no passed pawn logic) An important change for endgames, but cerainly not for the WAC suite. >> 2. correctly evaluate doubled pawns >> (previously, penalized 2 pawns on file, but ignored more than 2) Isn't this totally insignificant ? >> 3. correctly evaluate isolated pawns based on file >> (previously did not take file into account.) >> What's the correct evaluation of isolated pawns based on file ? Amir
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.