Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Let's have some fun!

Author: Richard G. Fadden

Date: 10:07:31 02/20/06

Go up one level in this thread


More postings from Vasik Rajlich

From: Re: Teaching computers to think
Posted by: Vasik Rajlich on March 15, 2004

< Skipping the comments of Andrew Wagner >

>>>>>>>>>>>>>> Cut/Paste of Posting >>>>>>>>>>>>>>
Hi Andy,

I'm only a few months ahead of you - so take this too with a grain of salt. At
first glance, selective search looks like a fairly easy thing to do. Obviously,
there must be some good ways to conclude that in some positions, passed pawn
pushes, checks, sacrifices, etc, are dangerous, and in others they are not.
Furthermore, it must be good to search good moves more deeply than bad moves.
(ie. Rac1 moving rook to an open file, vs Rca1 moving the rook back to a closed
file.)

However, it's not easy to make this work. I have tried a number of selective
search ideas and have not gotten very far. (By selective search I don't include
null move, and stuff near the leaves or under the leaves, ie q-search, I mean
fundamental selective search.) The only selective searching which I've found
which clearly improves my engine is check extensions. I think there are two
principle which work against selective search.

1) If you invest some # of nodes X to extend a line, then cutting the same # of
nodes from another line will cost you more depth in that line.
2) Each subsequent ply of search is more important than the previous one.

So, any selective searching has to be fundamentally correct enough to overcome
the burden of these two issues.

Of course, selective search if done right will work, the commercial programs
prove this. At the moment, I am completely re-doing my evaluation. I think
selective search must be tied to the evaluation, and in particular the part of
the evaluation which measures the high-value aspects of a position. In the
middlegame, this is king safety. In the endgame, this is passed pawns.

Re. the issue of how humans do it, I think it's not so important. If humans
could calculate 1M nodes/second, they'd be doing things differently anyway.

Good luck with Trueno.

Vas
<<<<<<<<<<<<<< End of Cut/Paste <<<<<<<<<<<<<<<<<<



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.