Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Question for the MTD(f) experts

Author: Vasik Rajlich

Date: 03:53:03 04/14/04

Go up one level in this thread


On April 14, 2004 at 06:10:55, Tord Romstad wrote:

>On April 14, 2004 at 05:48:07, Vasik Rajlich wrote:
>
>>On April 14, 2004 at 05:38:17, Tord Romstad wrote:
>>
>>>On April 14, 2004 at 05:32:37, Richard Pijl wrote:
>>>
>>>>>So my next question is, how do you normally populate a hash table with PV nodes,
>>>>>since we only get edge values during the search?  Do I need to follow the pv
>>>>>from hash to hash with a makemove for each succeeding pv node?
>>>>>{ICK}
>>>>
>>>>In addition to storing the move that gets a fail high, you could also store the
>>>>bestmove (i.e. score with highest value below alpha) in the hashtable. This only
>>>>makes some sense with fail soft. With fail hard (as TSCP is) you will (almost)
>>>>always get alpha as best value so the chosen move will be random.
>>>>
>>>>That way you will be able to construct an estimation of the pv, which will
>>>>improve with each iteration.
>>>
>>>Does this really work for you?  I once tried it, and the results were ugly.  In
>>>my experience, it is never a good idea to store a best move except when failing
>>>high.
>>>
>>>Tord
>>
>>This is a tiny but clear improvement for Rybka, and a huge improvement in PV
>>quality.
>
>Interesting.  It is probably time to experiment with this again.
>
>>How good is your fail-soft?
>
>It's probably awful.  I must admit that I have never understood most of the
>problems people talk about regarding fail-soft, and I have never given
>much thought to the matter.
>
>What are the characteristics of a good fail-soft?  Is there an easy way
>to measure it?

The better the fail-soft, the fewer nodes you search.

If putting fail-low moves into the hash table reduces your node count, your
fail-soft is good enough to support this.

>
>>I don't just mean returning the fail-soft value.
>>
>>I mean, dealing with lazy eval, dealing with scores returned from null move,
>>dealing with stopping the search in q-search, etc.
>
>When a fail-low or fail-high score is suspect for some reason, I usually
>return gamma-1 or gamma rather than the exact score (gamma is my search
>bound).
>

I can't think of a case where you'd want to do this.

>Tord



This page took 0.01 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.