[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] Pattern Matcher



How can things be dependant upon the bandwidth to RAM?

a) My processors here have a L2 cache so majority of lookups will
   come from that.
b) i guess you mean random access latency to RAM and not bandwidth.
c) it is a good idea to let those who know how to write searching software
decide how to use a pattern database.
d) I define them as positional, not as strategical. I know all beginners
confuse strategics with positional play. No big deal. Strategics is when
you play with a plan. Your thing at most can suggest which move to try first.


In theory it is possible to let your automatic recognized patterns auto
generate to C code. With evolutionary tree software you can build from that
a tree then which can hard code itself into C and just compile like an add
on C file. It can get evaluated then in a logarithmic way.

Patterns that require a black stone on h1, you can directly skip with such
hard coded knowledge.

Such automatic generated C code which automatically generates itself into a
very optimal decision tree can run very fast.

A big L2 cache would certainly help in such case to keep everything within L2.

At 17:46 8-11-2004 +0100, Frank de Groot wrote:
>> "At what speed can you predict for unknown position X the moves"
>
>It takes, on average, half a millisecond for a position.
>That includes the delta-administration of 2888 patterns and looking them up.
>
>It gets much faster when I exclude the largest patterns.
>The largest patterns are the size of the board (pat7) and 3/3 of the board
>(pat6).
>
>You can imagine the speedup when I start to exclude them after about 40
>moves.
>IIRC things become around 50 times faster.
>Things speed up again when I start to exclude the medium-size patterns at
>around move 150.
>
>Then, the time needed to pattern-evaluate a position is heavily dependent on
>the bandwidth to RAM.
>
>When I use only pat0 and pat1 (the smallest patterns), evaluation speed for
>a position is about 100 x the time needed to load a cache line.
>
>But I think the system should mainly be used for the strategic part of the
>opening, much less in search.
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/