Author: Colin Frayn
Date: 06:37:26 04/13/00
Go up one level in this thread
On April 13, 2000 at 07:08:12, Jan Pernicka wrote: >I would like to know if and how chess programs use randomness when > searching. Several points comes to speak about: > 1) choosing the move from openning book (it's not interesting even if you > are changing probability of choosing particular move) I think most programs use a weighted opening book with moves being more probable depending on how often they have been played successfully in the past. My program certainly does that, though I've been reducing the width quite a lot recently by setting the weight of a lot of moves to zero, just so that my program can learn a few variations extremely deeply and hopefully gain an advantage that way. > 2) choosing move when playing out of book (ie: you have several moves > with similar score - pick anyone...) Generally the fastest way is to retain the first one. That way you get more cutoffs at lower depths if moves are not good enough to _improve_ the best move. Removing the equality from beta cutoffs generally slows the program down noticeably. The advantage, as you mentioned, would be that you might avoid common mistakes, though the best way to do this, as you also say, is to just correct the opening book if necessary, or alter the weighting for your analysis function so that your program doesn't think bad moves are good. Often this is not so easy to do, but it would be the best option. Most of the blunders my program makes are because it doesn't search deeply enough and gets beaten on tactics nearly every time. This is quite annoying. I think that a small amount of randomness would only very rarely alter the move that my program would play in such positions, and often it would just alter it to another equally bad (or worse) move. > 3) (questionable) - to add little random number when evaluating position I used to do that when simulating lower skill levels, but I don't any more. It is quite slow, and it is also unnecessary. Sometimes your random number would correct for a bug in your analysis function, equally often it would make the analysis worse by moving it in the wrong direction. Cheers, Colin
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.