Author: Dann Corbit
Date: 00:36:24 12/08/98
Suppose that I have a database of millions of chess positions analyzed at long
time controls. Suppose further, that at some position "x", the analysis
suggests Nxh4 or some such. Now I go to the new position, and the current
stored analysis suggests Qxh4. This process continues, and I update the
original pv as we go along, and also the ce to reflect current standings.
After this process has been run to a termination {mate/draw/*}, we process each
of the alternate moves instead of Nxh4 from "x". Perhaps after each of these
has been analyzed, Nxh4 no longer looks best. We can repeat this process as
many times as we like. If some move is suggested but never taken, we simply add
it to the database by calculation. Since we are only examining the suggested
move from a given point, and following to the end, is seems that substantial
refinement could be achieved after a few cycles of the system. Even when we
choose an alternate, we only pick the move that looks best from the alternate
from an "annealed ce" standpoint.
We could try other annealing methods also. Start from the terminal positions
and work backwards. Trace over the best winning percentage each time. Lots of
interesting alternatives. It seems that a method such as this could enable us
to look *much* farther ahead than traditional methods.
When a new move is discovered that "clobbers" our discovered notions, we only
have to heal only for inputs and outputs from that node.
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.