Author: Terry Presgrove
Date: 00:34:58 08/18/99
The following is a response by Shep to a question I proposed
regarding at what % do current top programs anticipate/predict their
opponents next move?
"During those games I mostly play, namely top professional vs. top
professional, the percentage is typically 50-75%.
(If one opponent is significantly weaker or both are amateurs, the quota
can be as low as 15-25%.)
In more "standard" and "ordinary" middle games, it can be up to 90% if the
programs do not have a very particular style (like Chessmaster or CSTal).
The current game between Shredder 3 and Hiarcs 7.32 had about 85-90% except
for the phase right after the opening.
Some programs predict very well while themselves being very unpredictable
(e.g. Tiger) or vice versa (e.g. MChess).
Of course these numbers and observation are only exact within a certain
margin since I don't keep precise statistics about this.
But it is a good indication that
a) it would have a significant impact to be playing without permanent brain
and
b) differences in playing strength arise just from these seemingly few
cases where the programs still disagree.
As an aside, but showing the same thing, I have been using my programs for
post-game analysis of my correspondence games and typically see one of the
following two behaviours:
1) all programs except 1 or 2 suggest the same move
2) the programs are almost equally distributed between two choices
(These data are reliable since I have post-analyzed about 30 games of 25-85
moves each with 5-10 programs at 10 min/move. Now that I think of it, I
will probably put this up on my site as well.)"
There are many points to made from the preceeding I would make the
following:
(A). If program A predicts 75% of Program B moves but program B only predicts
50% then we would have a real probability that the game would be influenced
significantly by ponder being off.
(B). If the difference between programs playing strength stems from those few
times when they disagree, what does this say about the differences between
the top programs and the various programming techniques?
(C). It may be that this sort of data would help us in understanding the
complexity of matching programs on same machine with ponder off. It is
presented merely has a sounding board for those of you who are the experts
in this field.
Thanks to Shep
TP
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.