Computer Chess Club Archives


Search

Terms

Messages

Subject: Search behavior in a case of root fail high/low

Author: Sergei S. Markoff

Date: 04:56:17 03/10/04


Hello!

I think that there are a lot of ideas that can be used to reduce search tree,
clever changing a window size of a root search.

There are some weaknesses of MTD(f) or MTD(f)-like schemes that not allow to use
this schemes by the most of engines. The first is very hard depending of hash
storing/probing efficiency; the second - troubles with null-move. I tried
something like MTD(f) before, including trying to determine move singularity in
position (actually we don't need the exact score of move sometimes to make best
move).

Anyway there are some ideas how to improve root search. You can read Ed's work
about Rebel "root-behavior" it seems to be interesting.

SmarThink 0.17 uses for the first move alpha=eval-1, beta=eval+1. There are at
least three frequent cases where it returns eval:

1) Draw
2) Checkmate
3) The last best move in PV was already found by quiescence during previous
iteration.

If this search fails low or high SmarThink enlarges window by 0.15. I'm not sure
that it's the best value. If this search fails also than window will be enlarged
to INFINITY.

Now I'm trying other scheme. After 0.15 window fails I'm first trying to enlarge
it by 1.50. But there are one more ideas. If eval fails high (eval<alpha-1.50)
than may be better to stop searching this move and try to get alternative?

I read in some publications that It's a good idea to reset depth to the starting
when finding an alternative to failed move. Is Rebel the single engine that uses
it in practice?

Is it a good idea to uses some knowledge to choose windows for a root search?

Best wishes,
Sergei



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.