Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: search speed

Author: Stuart Cracraft

Date: 16:32:31 06/25/04

Go up one level in this thread


On June 25, 2004 at 19:20:08, Dann Corbit wrote:

>On June 25, 2004 at 19:11:34, Stuart Cracraft wrote:
>
>>On June 25, 2004 at 17:46:37, Dann Corbit wrote:
>>
>>>On June 25, 2004 at 16:55:47, Stuart Cracraft wrote:
>>>
>>>>On June 25, 2004 at 16:46:31, Dann Corbit wrote:
>>>>
>>>>>What is your branching factor?
>>>>>
>>>>>The best programs have a branching factor between 2 and 3.
>>>>>
>>>>>If you already have a branching factor in that range, do not expect any more
>>>>>really dramatic speedups.
>>>>
>>>>This seems like a very useful measure to implement but I am not
>>>>sure how it might be done.
>>>>
>>>>Take the # of total legal moves searched at each node (prior
>>>>to cut) and divide by the total number of nodes searched across
>>>>all nodes?
>>>>
>>>>Time between iterations in mine goes up about 2x
>>>>to 6x per iteration but this is obviously a different measure
>>>>than what you are suggesting.
>>>
>>>That is close enough.  Look at the deeper plies to see what sort of time ratio
>>>you are seeing.
>>>
>>>If those average around 2-3, there is not a huge amount of fat to trim.  If they
>>>average around 5-6, then you will still see some huge speedups.
>>>
>>>Are you using a pvs search?  That cuts a lot of nodes.
>>>
>>>IID gives better move ordering and sometimes cuts nodes because of that.
>>>
>>
>>What is IID? I have iterative deepening...
>
>Internal iterative deepening.
>http://www.chessbrain.net/beowulf/theory.html
>
>>>Without knowing what techniques you are using, it is hard to speculate about
>>>what sort of speedups you might see.
>>
>>Yes I use PVS though currently the pv is not searched first and I wonder
>>how much that will help. However, the best move found by the previous
>>iteration is searched first on the next iteration as a very
>>cheap approximation. Eventually I will walk the transposition table for
>>the PV.
>>
>>Briefly, it has:
>>o iterative deepening (hash table not cleared through iterations)
>>o history heuristic (thank you Jonathan)
>>o 1-tier 1M entry transposition table (algorithm: replace-always)
>>o null move with reduction factor R=3
>>o quiescence search all caps MVV/LVA (no SEE though)
>
>SEE will probably help a lot.

Yes -- I must.

>
>>o extensions: move in capture search to get out of check
>>o evaluation - just pc/sq at the present.
>>
>>There are no other search extensions, no fractional, extensions, no
>>futility, no razoring.
>>
>>So that's it. Bare bones. Just looking for big things I've missed
>>that won't take an arm and a leg to implement before the real
>>research can begin. Looking for double-digit improvement possibilities
>>that remain, if any.
>
>What is your fail high percentage?  If it is below 90, then you can have a large
>improvement in fewer nodes searched by better move ordering.

I need to collect this statistic.

>
>>I wonder if searching the PV first instead of just the PV's first move
>>first would make a big difference. Probably so. Can't think of much
>>else on my own presently.
>
>I assume that you use an aspriation window?

Yes -- alpha=eval-1p to beta=eval+1p of previous iteration's search value.
If 1st iteration then eval is set to static board eval without search.

>
>It looks like most of what is left for improvment could be through razoring or
>other frontier pruning types of ideas.  But they have danger, of course.
>
>What is your NULL MOVE set to for reductions?

It is R=3. Have tried 2, 4, 5, ply/2 and ply-2 where ply is the current
iteration depth being searched. Every higher search results in a massive
additional decrease in total nodes searched. I haven't played out any
games though so I don't know the quality of the moves with the much
higher reduction settings for the null move search. I question the
R=2 and R=3 recommended settings for high-depth searches (see other
thread.)

Thanks.

Stuart



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.