Re: [computer-go] Computer Go hardware

[Paul Pogonyshev <pogonyshev@xxxxxxx>]

Frank de Groot wrote:
> > One problem with using all possible resources is that improving and tuning
> > program becomes more difficult (as it is slower.)
> 
> 
> Not at all.
> RAM means "Random Access Memory".
> Which means that any random location in RAM is accessable in approx. the
> same time.
> 
> So if you have a program that needs to lookup 1000 patterns in a
> 10,000-pattern database or a program that needs to lookup 1000 patterns in a
> 100,000,000-pattern database, this takes *exactly* the same time.

You misunderstood me.  Let's say program A normally plays a game in 10
minutes.  So, you let it play itself 100 games and then look through the
game records in search for repetitive mistakes.  Fine, you find some and
tune them.  Now suppose you have another program, B, which normally plays
a game in 3 hours.  Just playing 100 games with it will take several days.

Speed also matters when you make tuning changes.  They are usually sort
of incremental: you try and fail, try again and again, until you find the
right (or at least a working) way to solve your problem.  The longer your
program "tests" your changes by solving some regression problems, the more
time (and your patience) is spent.

Even if your program can easily adjust its speed, i.e. be fast enough
when you are at tuning, there are still problems.  Scaling _does_ change
program strength and it can hide/reveal mistakes.  If your program thinks
for 30 minutes on a move and makes a mistake, it can think for 20 seconds
and play a tesuji, by luck, because of horizon effects, etc.

Paul

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