Re: [gnugo-devel] pre-rotate implemented

[Dave Denholm]

Trevor Morris <address@hidden> writes:

> http://www.public32.com/games/go/trevor_3_11.2
> 
>  - pre-rotated pattern generation in mkpat. (-p option) implemented.
>  - %2m added as valid syntax for gprintf
>  - relabel rotation matrices in transform.c
> 
> In transform.c, "invert" didn't make sense to me, so I attempted to
> clarify the meaning of the matrices.
> 


> This patch causes no changes, as no pattern databases use
> the '-p' option yet.  I was surprised that there wasn't a
> large performance increase.  I was also surprised that the
> database size impact is enormous.  Don't even try this
> option with the owl_attack or owl_defend datbases.
> 

What kind of performance difference did you see ?

I'm still interested in seeing what happens if we let the dfa
provide a shortlist of patterns which *might* match, and let
a separate matcher check them in detail. This might help with
the explosion in data size.

And of course I'm still trying to see what the impact of storing
one complete list of patterns in each transition to error state.
I think I'm almost there...

I'm not too surprised by the database increase -
Just out of interest, I tried kludging mkpat to add all 8 transforms
of a pattern to a dfa, to see how well my implementation scales.
(Adding one pattern at a time in situ).

Unfortunately, I got fed up after 11 minutes on owl_defendpat.db :-(
The current 3.3.8 mkpat did it in about 3 minutes, but it consumed
vast quantities of memory (100Mb ?) to do so.



Maybe building dfa's at run time is something we could do during
opponents thinking time..? Should be plenty of time before
the position gets interesting enough to need them.

( Unfortunately, could be hard to debug, since the bug could be
  due to a dfa which was built dynamically, and not available
  when reproducing.
)


dd
-- 
address@hidden          http://www.insignia.com