Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: How many nodes?

Author: margolies,marc

Date: 14:42:09 11/09/03

Go up one level in this thread


Thank you for your response,Scott.
When I study a single position aa EPD, I usually go at least 100,million nodes
sometimes as high as 350,million which is roughly 17 or 18 ply.
What I do basically is watch for changes in the first two moves of both sides in
the pv. when at this depth  the pv.completely stabilizes then I an satisfied my
answer is accurate.
Regarding the specifices of your response, the choice at 10,thousand nodes is
very stable, but changes forcing new evaluations then occur at 300-400,thousand
nodes, which I prefer to analyze to some clarity.
I suppose that were I examining later stage middlegame positions, your
recommendations may be truer to my results.
Sometimes I read a game between strong players and 'see a better move.' Since I
do not believe my self about this, I let deep analysis help my resolution. And,
not surprisingly, the longer I run an engine, the better the master's eval looks
but often it looks very poor at start-- and sometimes it stays poor for a long
time.




On November 09, 2003 at 07:49:15, scott farrell wrote:

>On November 06, 2003 at 00:56:09, margolies,marc wrote:
>
>It all depends in the accuracy you are after. It will be accurate after only a
>few hundred nodes, unless there are come specific tactical moves.
>
>Most modern engines find simple tactics in a few thousand nodes.
>
>In a million nodes most obvious tactics are overcome.
>
>If an engine cant see a tactic in 10 mill nodes, it is unlikely to find it.
>
>If you are analysing full games, use a small number of nodes, and if the score
>jumps, keep the hashtable and go back and redo a few nodes deeper to see the
>earliest point you can pick up the tactic.
>
>Scott
>
>>When examining a position deeply with an engine, say for the purposes of placing
>>the evaluation of the position within a tree of candidate variations, how many
>>millions of nodes of testing are necessary in order to get the closest
>>approximation to that engines  best numerical assessment of a chess position and
>>its roughly corresponding possible variation move sequence?
>>Is there a numerical methods heuristic, or some other rule of thumb which I
>>might chose to use to attack the certainty issue? Maybe a statistical
>>regression? Should it be engine specific (and correlative to the kinds of
>>positions under assessment) or generic?
>>I welcome any suggestions here. Thanks-Marc



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.