Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Transposition table hit rate?

Author: Roberto Waldteufel

Date: 04:33:03 12/21/98

Go up one level in this thread



On December 21, 1998 at 05:48:44, Pauli Misikangas wrote:

>Hi all!
>
>What would be an usual transposition table hits/lookups rate when using PVS and
>iterative deepening? How many of those hits (the position was found from the
>table) should cause a cutoff? I'm asking this because I suspect that my
>transposition table is not working properly. My hit rate is about 20% and only
>about 8% of the lookups cause a cutoff. Should I be worried?
>
>Merry Christmas!
>
>Pauli Misikangas

Hi Pauli,

I think both your hits/probes ratio and your cut-offs/probes will depend very
heavily on the type of position you have at the root. Of course the size of the
table is also an important factor, but I would suggest that you might find a
much higher ratio for endgame positions than you do in opening and middlegame
positions. The fewer pieces there are, the more transpositions there are and so
the more hits you get. Try it on something like a king and pawn endgame where
most of the pawns are blocked - the famous Fine #70 springs to mind as an
excellent example. You should get about 90% hits or more in that sort of
position.

As for the cutoffs, don't forget that with iterative deepening you very ofter
reach a stored position with increased depth, so you cannot use the hashed value
or bound (so no cut-off), but that does not mean that the probe was wasted
effort, because you can still use the hashed best move and search it first. The
improvement in move ordering should at least pay for the overhead of the table
look-up

One more point about hash tables - take great care how you set up the table.
Make your hash records of a size to fit well into the cache lines (32 bytes), ie
ideal hash record sizes are 8,16 or 32 bytes in length on a PC, and make sure
your table is aligned by 32. I found that not doing so slows down the program by
about 15% although this figure may be different depending on your
implementation.

I hope the above is useful. Persevere! It WILL be worth it when you get it all
in sync. Good luck.

Best wishes,
Roberto



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.