Re: [computer-go] Pattern database

[Paul Pogonyshev <pogonyshev@xxxxxxx>]

Frank de Groot wrote:

> Is there some data on pattern matching in Gnu Go?
> (how many patterns used,

Number of patterns used in life-and-death reading (all matched
with DFA matcher): over 900.

> pattern shapes,

Don't understand this.

> speed of matching as a function of number/size of patterns,

All patterns are matched at once (actually they are separated in
three databases, but only because they are used for different
purposes.)  So, matching time depends on number of patterns
indirectly: because the database size increases, there are more
memory cache misses and there are less early cutoffs.  If you
ignore these factors then it is O(1).

As long as size of patterns is considered, then speed of match
is approximately O(N) where N is the number of intersections in
an average pattern.

> how they are harvested,

Manually.

> how much they contribute to the playing strength and how they
> contribute to playing strength?)

GNU Go depends on patterns matched with DFA completely.  Other
patterns are important, but not vital.  Exact impact on playing
strength is not known (to me.)  Don't know if it was ever
measured.

DFA-matched patterns are used in life-and-death reading for
move suggestion.  They are actually the primary source of moves
in this reading for GNU Go.

Paul

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/