Re: [computer-go] SlugGo v.s. Many Faces update

[Evan Daniel <evanbd@xxxxxxxxx>]

On Thu, 31 Mar 2005 18:02:56 -0800, David G Doshay <ddoshay@xxxxxxxxxxxxxxxxx> wrote:
> 
> On 31, Mar 2005, at 4:38 PM, Evan Daniel wrote:
> >
> > It's starting to sound more and more like GoFigure :)  (With the
> > exceptions that GoFigure isn't clusterable and only plays 9x9, and
> > searches to the end instead of a fixed depth...)
> 
> In what ways?

Mostly in that you're now doing branching at depth > 0.  I've also
done some experiments with varied branch factors, eg have branch
factor > 1 while on the main line and = 1 off the main line; that
appears to be a better use of time, but I don't have actual data to
back that up.  What with my alleged plans to work on clustering
support for GoFigure, and eventually 19x19, the two might end up
looking even more similar :)

> 
> > Have you solved the problems you were having with poor opening play (I
> > have my own answers to that question, and would be curious to
> > compare)?
> 
> We have a kludge I am not proud of. I hope to improve upon it soon.
> We have been busy trying to make up for all the lost time rather than
> trying to change features that may be ugly but work well enough for
> now.

My solution was to build an opening book :)

It has its pros and cons; I can talk about it more if people want to
hear details.

> 
> > What branch factor settings are you using,
> 
> We are presently doing combinations that end up with 24 paths:
> 24 - 0
> 12 - 2
> 8 - 3
> 6 - 4
> 6 - 2 - 2

Interesting.  How useful do you find the lower priority branches to
be?  (The easiest definition I know of is % time chosen -- I haven't
examined it very recently, but my recollection was that showed a steep
drop off in GoFigure, suggesting that the lower priority branches
aren't as time efficient).

Of course, with alpha beta pruning, those branch settings don't really
produce 24 paths in some sense ;)

> 
> We have not tested them all yet. We will look at those results
> before deciding about trying combinations that start with 4
> (and would allow for richer branching after that). We have
> decided against branching like 12 - 0 - 2 (sub-branching
> only on our moves).
> 
> > and have you noticed much strength variance?
> 
> "Much strength variance" is what I notice most clearly,
> even with all the parameters exactly the same! ;^)

Got it ;)

I think I'd see the same thing, except that programs tend to play in a
more reproducible manner on 9x9, so if I run a significant number of
games I get duplicate games and the results converge ;)

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