RE: [computer-go] Tactical reading

["Mark Boon" <tesujisoftware@xxxxxxxxx>]

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of David Fotland
> Sent: Saturday, December 11, 2004 23:37
> To: 'computer-go'
> Subject: RE: [computer-go] Learning : was Chess programs versus go
> programs
>
>
>
> Mark,
>
> I tried limiting tactical search to 3 liberties instead of 4 on a set of
> life and death problems.
>
> On about 1000 problems, it solved 9 fewer, and spent about half the time
> getting the solutions.  I expect some evaluation tuning will easily get
> those 9 solutions back.

So with a simple change you get a two-fold speedup! That is good news. And
possibly your tactical search was hiding one or more flaws in your
evaluation, which when fixed will mean some general improvement as well.

Reducing tactical search saves half the time on L&D reading in Many Faces?
Maybe now you understand why in L&D my getNlib routine, which is the base of
my tactical reader, was using 30% or more of the total time. Using the new
chain-administration module, that incrementally keeps stone-lists, liberty
lists and neighbours instead of the tailor-made code that uses getNlib
didn't speed things up. It simplified things, as it reduced code by some 350
lines, but runs significantly slower.

This chain-administration module was made in a day though, where the
tailor-made tactical reader must have at least a months worth of
optimization work in it. So maybe the chain-administration will just need
some more optimizing to become a viable and simpler alternative.

> So I think you are correct that using 4 liberties for tactical
> stability is
> better than using 5 liberties once you have good life and death
> evaluation.
> The next version of Many Faces will have this change.
>
> David

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