RE: [computer-go] Designing faster, better influence functions

["Mark Boon" <tesuji@xxxxxxxxx>]

Would you care to explain? I'm not familiar with the terms MMX,
morphological kernels or Manhatten distance.

The faster you can make it, the better. But influence itself doesn't really
have that much value. It's what you do with it. For me board evaluation
starts with influence. But then a whole lot of other stuff needs to be done
before you can come to a score or another kind of result, like status of a
group. With current hardware I don't know (yet) how to do more than a few
thousand moves per second. Unless I find other uses for the influence
function, it's called once or twice per move. Making the influence function
an order of magnitude faster saves me at most a few percent. Only worth it
if it doesn't compromise the quality. And is not too hard to obtain. Two
orders of magnitude faster is not going to be noticable. But maybe once you
have it you'll see uses for it that are currently not so obvious.

By the way, you'll either need to bounce off the edge of the board or you'll
need to compensate for the extra strength that proximity to the edge
provides. I never bounced off the edge, although I think it would give
better results. Instead I use lower tresholds for the 1st and 2nd line. I
should maybe see if I could gain speed or simpler code elsewhere if I would
spend a bit more time caclulating better influence at the edge.

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Frank de Groot
> Sent: Sunday, October 31, 2004 18:56
> To: computer-go
> Subject: Re: [computer-go] Designing faster, better influence functions
>
>
> From: "David Fotland" <fotland@xxxxxxxxxxxxxxxxx>
> Subject: RE: [computer-go] Designing faster, better influence functions
>
>
> > I can't find the detailed description.  The basic idea is:
>
>
> Thanks for this.
> I have discarded my own function in favor for a variation on your
> function,
> as my function did not stand up to testing.
> As soon as I left Bouzy's example, results were poor.
>
> Today, I designed a non-iterative method to your influence function that
> should take (in case of MMX, much) less than 1 microsecond for a
> delta-update. There appears to be a method that uses magic with
> morphological kernels instead of iteratively looking around.
>
> The only catch is that "bouncing off the board" is not
> implemented, and that
> the maximum sphere of influence is 6 expansions. I managed to keep
> aritmethic Byte-based.
>
> I wonder why you let the function bounce off the board, and
> whether it would
> result in a substantial loss of usability of the function. The same for a
> sphere of influence with a Manhattan distance of about 11, as you seem to
> use.
>
> I would like a <= 100 nanosecond delta-update speed so it would
> be great to
> stick with this limited design.
>
> _______________________________________________
> 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/