Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: triangular pv vs. hash move pv

Author: Robert Hyatt

Date: 07:12:01 09/12/04

Go up one level in this thread


On September 12, 2004 at 00:56:28, Michael Henderson wrote:

>On September 12, 2004 at 00:25:58, Robert Hyatt wrote:
>
>>On September 11, 2004 at 11:17:35, Charles Roberson wrote:
>>
>>>
>>>  This discussion brings out an interesting thought. I agree with all of
>>>  Bob's statements, but how about a change of use?
>>>
>>>   Lets assume a single stage cache and replacement scheme of
>>>   current depth > previous depth. Also, we have a collision test to ensure
>>>   we really are on the same position.
>>
>>If you do simple depth-preferred you will crush performance, as your table will
>>fill with deep draft entries and local sub-trees can't store anything at all..
>>That is a known problem and is the primary reason for the two-tier Belle
>>approach many use...
>>
>>
>>>
>>>   Now, the PV is set but in a later tree branch one of the PV positions
>>>   is overwritten due to greater search. True this changes the PV from what
>>>   you originally setup but there may be a positive side effect of this.
>>>
>>>   If we only allow the last few moves of the PV to be changed by hash
>>>  replacement(I can think of two ways to do this) then wouldn't we get better
>>>  moves for the PV tail? Thus, improve move ordering for the next iteration due
>>>  to a somewhat better PV?
>>>
>>>  Charles
>>
>>Remember that two different positions can map to the same table address, even
>>though the hash signature is different.  Short PV = bad move ordering.  Also if
>>you do only depth-preferred hopefully you will prefer exact over upper or lower
>>entries, and that leads to lots of problems with the wrong PV move that comes
>>from a _shallower_ search than the real PV move...
>
>Does this mean that the "shallow" entries lead to complete blunders in the
>extracted PV?  Or is that caused by something else, like no path info in the
>hash table?  I see a lot of chances for bad moves...

The shallow entries could lead to most anything of course.  Blunders are far
less likely but if the search terminates poorly when time runs out, it is even
possible that illegal moves get stuck in the hash as the search unwinds and
returns to the root position.




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.