Author: Tord Romstad
Date: 11:14:51 06/17/04
Go up one level in this thread
On June 17, 2004 at 13:49:46, Robert Allgeuer wrote: >What is the "blemish effect"? Is it a jump in the score? I don't know (I haven't read Berliner's paper), but I guess so. >Why is it actually bad to have discontinuity in the scoring function? There are at least two important reasons for this. The first, and most important, is that the program will often behave very strangely in positions where the material configuration is close to such a discontinuity. Assume that you have a hard limit between the middlegame and the endgame, and that you use different evaluation functions depending on the game phase, with no interpolation between. Because neither your endgame eval nor your middlegame eval will ever be anywhere near perfect, there will often be positions where the middlegame eval overevaluates the position compared to the endgame eval, or vice versa. This can cause the program to try to avoid exchanges at all costs, for no good reason. The second reason is that the search algorithms used by chess engines are most efficient when the scores don't vary too much. Wild jumps in the score when the border between the middlegame and endgame is crossed would make the search less efficient. From a less technical point of view, you could also argue that the border between the middlegame and the endgame in chess really *is* fluid. In a typical chess game, there isn't a sudden moment when the king no longer n needs to hide near the corner and should start moving towards the centre. A more usual scenario is that the importance of keeping the king safe gradually decreases when pieces are exchanged, and at the same time the importance of having an active king increases. The most simple and obvious way to implement knowledge like this in a chess program is to interpolate between a middlegame and an endgame eval. Tord
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.