Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The Dangers of Positivism with Numbers

Author: Ratko V Tomic

Date: 06:51:13 11/02/03

Go up one level in this thread


>  I hold up the chess truth that more depth will
> always be advantageous.
  ^^^^^^

Problems with "always":

1. That depends on what is being traded off for
   the depth gain. If the quality of the final
   evaluation is sacrificed too much, the strength
   may decrease when you raise depth and evaluate
   more poorly. E.g. Hiarcs often plays better at
   depth 1 (i.e. it plays with pure high quality
   root evaluator) than many other programs with
   depth 3-4. So if Hiarcs were to drop its quality
   of evaluation to increase the depth to match the
   depth of the simpler programs, its strength would drop.

2. In addition to evaluation quality/depth trade-ff,
   merely showing that program changed its mind in
   1.2% cases when increasing the depth from 20 to 21
   plies, doesn't mean that the new move was stronger
   in all 1.2% of cases (e.g. just add the percentages to
   say, depth 5, and you are already to 118%, which means
   some moves get discarded then picked again at greater
   depth, hence some depth increases had picked worse
   choice than their shallower predecessors).

3. It is conceivable that the increase of the "evaluation noise"
   with greater depth may at some depth _range_ make a particular
   program play worse than the same program at lower depths.
   (Obviously, this negative gain range would eventually turn
   around to positive since the further increase in depth can
   reach the true, or at least high accuracy, terminal positions
   in sufficient percentages to offset the increased noise
   in the remaining cases.)






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.