Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Yet another 'horizon effect' problem

Author: h.g.muller

Date: 02:58:04 02/15/06

Go up one level in this thread


This is why I am experimenting with double accounting of ply depth, one counter
for each side. Delaying tactics, like capturing the Rook, then only burn the ply
budget of the side that initiates it if the recapture is a (singular) extension,
and you will never be able to push anything over the horizon.

But you have to be careful to also award extensions for solving threats like P
attacks Q, or check evasion, which is much more difficult to control because
threats can usually be solved in many more ways than recaptures.

To recognize threats like forks or attack on trapped pieces at the horizon, you
will have to judge quiescence from both sides, i.e. if you don't have any
QS-allowed moves yourself, not automatically rely on static eval but throw in a
null-move first to make sure there is no threat. And if there is a threat, award
yourself the extension needed to neutralize it. (Which for forks of course is
not possible, so you recognize the loss.) To prevent search explosion there
should be strict limits on what the threat-evasion can do, it should not be used
as an excuse to start a new offensive tactical combination on its own (that is a
strict nono beyond the horizon, since it can be done from almost any quiet
state). To achieve this, reduce beta to the current eval, just to prove that the
node is not worse than that, i.e. the threat can be dealt with. (If your
evaluation contains bonus points for having the move, you could negate these in
a situation where there is a threat that you can solve, because apparently your
move is already spoken for.) And of course you don't search just any move, but
only the standard set of solutions to a threat (i.e. retreat, capture the
attacker, block the attack, defend).



This page took 0.01 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.