RE: [computer-go] Modern brute force search in go

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Vincent
> Diepeveen
> Sent: Sunday, November 07, 2004 18:56
> To: computer-go; computer-go
> Subject: RE: [computer-go] Modern brute force search in go
>
> I will use 1 million nps now.
> That's 10^6. Important moves where score drops you time extend. Not sure
> what time you can time extend. This world champs Shredder searched at 1
> important move against me for 60 minutes nearly. Say 10^3.
> 10^14 to go.
> log(10^14) / log(2) = 46.5
> That's 46.5 doublings in hardware needed to get 10^14 faster.
> With 18 months for the original Moore law, that's
> 46.5 * 18 months = 70 years.
>
> That's a lowerbound of 70 year to go.
>
> It can get better, not worse.
>

Hmmm... 70 more years of Moore's law? As much as I had trouble imagining
what would be possible with a computer a 1000 times faster than the one I
had 15 years ago, 70 more years is really beyond my grasp. And still today,
with current hardware, some great improvements are possible. I think only a
small percentage of the real potential of current hardware is used by Go
software.

A computer 100,000,000,000,000 times more powerful? (Sorry, I had to type
out the zeros to get the idea.) Yes, I think there's a definite chance you
can beat a pro with that kind of computing power to a pulp.

Even 30 years, or a computer that is millions of times faster, is hard to
imagine. And we can't use that power for evaluation? 1 million whole-board
evaluations per second is already not achievable, let alone when we want to
make evaluation more sophisticated. Which will be needed to get it to the
level required.

To have Moore's law continue to be valid for that long seems very optimistic
to me. In the meantime I think alternative approaches will have a better
chance than brute-force.  Maybe by then, with that much computer-power, we
can simulate a complete brain :-)

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