Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Experiment: Search depth and ratings

Author: James Swafford

Date: 10:20:53 07/31/02

Go up one level in this thread


On July 31, 2002 at 02:14:50, Russell Reagan wrote:



>Has anyone done any experiments where they took a simple program and computed
>ratings for different levels of search? I was thinking it might be interesting
>to do this experiment, but I figured if someone has done this before it would
>save me the trouble.
>
>I was thinking about creating a simple alpha-beta engine with mainly material
>evaluation, and perhaps some other small things like a piece-square table, and
>maybe a bonus for castling. That would seem to help the program do things that
>even beginners do, if for no other reason than they were taught to "control the
>center" and "castle early". The main reason for adding those two things would be
>so the program wouldn't be deciding on 1. h3 just because it was the first move
>searched and the material evaluations all came up even.
>
>My goal here is to compute ratings for various depths of search when using
>(basically) material only evaluation. I would like to know, for example, how far
>you could expect to get as a human player if you were able to catch all tactics
>and combinations at a depth of 1, 2, 3, 4, and so on. I have heard people say
>that a human can get to expert level (2000) by mastering tactics, and I would
>guess they also know enough positional knowledge to get by.
>
>If anyone has done anything like this before, I'd love to hear the results. It
>would save me the time of doing it myself, although I might do it anyway just
>for kicks.

Yes... I don't have time to read the other posts (I'm in a conference hall;
they're only giving me ten minutes!), but Ken Thompson did this in the
early eighties.

I believe most computer chess programmers will tell you that there doesn't
appear to be any diminishing returns in computer chess, but Schaeffer
did some interesting work showing that actually, there are.  There is a
catch to this, though - there are diminishing returns _if_ you limit the
games in length.  (Say, you adjudicate at move 40 -- 8 ply players will
do better vs. 9 ply players than 3 ply players vs. 4 ply players).  This
doesn't show up much _over the board_, though, because as search depth
increases, the game lengths become correspondingly longer.  As the games
get longer, there is increased chance that the player with greater search
depth will find a better move.

Shoot... only six or seven more minutes. :)

--
James


>
>Thanks,
>Russell



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.