Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Iterative deepening -- Why add exactly one ply?

Author: Vasik Rajlich

Date: 08:55:44 05/27/04

Go up one level in this thread


On May 27, 2004 at 06:46:50, Uri Blass wrote:

>On May 27, 2004 at 06:39:23, Vasik Rajlich wrote:
>
>>On May 26, 2004 at 13:49:38, Tord Romstad wrote:
>>
>>>On May 26, 2004 at 13:34:23, Robert Hyatt wrote:
>>>
>>>>I have used "IID" for years, but in a very restricted way, namely to handle the
>>>>case along the PV where I have no hash move.  I've never tried it _everywhere_
>>>>before, so have no data.  But I intend to try to see if it is something that
>>>>could work, or if it is a waste...
>>>
>>>I am fairly sure you will find that _everywhere_ is a waste.  It is probably
>>>not worth doing near the leaf, you have a hash table move to search, or when
>>>a fail-low is most likely.  Perhaps you should also use a somewhat bigger
>>>reduction factor than in your "along-the PV IID".
>>>
>>>Note that it could also be interesting to look for good ways to make use of the
>>>return value of the internal search.  It gives a reasonably reliable estimate
>>>of the value of a full-depth search, and can be useful as an ingredient in
>>>pruning tricks.  The most obvious (and entirely risk-free) case is when the
>>>reduced-depth search returns a mate score.  When this happens, it is clearly
>>>not necessary to do a full-depth search.
>>>
>>>Tord
>>
>>Yes, there is lots of room for playing with IID.
>>
>>Note that 95% of all nodes fail high in some way, so you can be pretty
>>aggressive.
>
>The % of all nodes that fail high is dependent on the engine but it does not
>seem to me logical that 95% of all nodes fail high in some way.
>
>After every nodes that fail high you may have many nodes that fail low because a
>node that fail high means that all the node one ply after it must fail low.
>
>Uri

Here are some sample fail high %s, from Rybka's last run:

93.5
93.6
92.9
93.4
93.0
94.4

My search has three layers, these are from the top layer.

I think the #s would be slightly lower for a PVS search instead of MTD (f).

As for the "logic" - don't forget that fail-low nodes will spawn more children
than fail-high nodes.

Vas



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.