Author: Robert Hyatt
Date: 08:42:38 02/10/04
Go up one level in this thread
On February 09, 2004 at 14:06:00, Pham Hong Nguyen wrote: >>There is a danger. You then mix perfect information with non-perfect >>information. And inconsistent behavior is the result. IE if I probe everything >>up to the q-search, but then only probe 3-4 piece tables, it is more than >>possible that I pass over a drawn 5-piece ending, to reach a 4-piece ending I am >>winning or losing. Comparing the scores then becomes a problem and alpha/beta >>doesn't like doing that very well. It is bad enough to have some artificial >>boundary beyond which you don't probe, as that allows for the classic horizon >>effect where you push a capture off to beyond where you see that it loses or >>draws according to the table, then you can make your plans accordingly, only to >>realize in a couple of moves that you are committed to a path that you thought >>was winning, but isn't. Adding another discontinuity only increases the chance >>that such horizon effects happen. > >I see the danger. However, I think that is the matter of programing and setting >parameters, do you think so? It depends on your goal. If you suddenly turn on or off critical pieces of evaluation, at arbitrary depths in the tree, then you have a discontinuity at that point. That leads to some strange problems at times, as the program can fiddle with things to make the discontinuity occur at a point of its choosing. For example, I am black and you are threatening to take my pawn on f7. I might move my king to g8 so that Bxf7 is now a check, driving the search 1 ply deeper, to let me see something just beyond that discontinuity. Or I might move it to f8 so that Bxf7 is not a check, reducing the search by some amount without the check extension being used, to reduce the depth and where the discontinuity occurs in the tree. A full-width search is a _very_ strange animal in such circumstances, as besides just searching and evaluating endpoints, the search is fiddling with things to try to control the depth at which something occurs, so that the discontinuity makes the result look better than it should... or worse for the opponent, etc. > >I love to see that Crafty is always the most advanced open source code and takes >the highest titles again :) > >+Pham
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.