On Fri, 2007-01-19 at 12:27 +0200, Petri Pitkanen wrote:
> But also in chess I think there is a limit on what can be gained by
> adding time. After so and so many minutes/move your decision would
> not be any better.
This is not true in the positions that matter. Again it's a human
time perception issue. Have you ever watched a computer chess
program solve a difficult problem? It goes like this:
1. At first it changes it's mind a LOT - display flashes a lot.
2. Then it slowly settles down - perhaps changing after a few
seconds.
3. Then it seems to like just one move (the wrong one) and it
might spend hours depending on how difficult the problem is.
But we think of time in a linear way. After a few seconds of
watching a display we get bored. If it takes 10 minutes to
switch moves it seems like forever.
The difference in strength from 1 second to 1 minute is ENORMOUS.
We don't mind that because 1 minute we can live with. But if
it has been thinking for 1 hour, an extra minute doesn't give
another enormous increase, it perhaps adds 1 ELO point, virtually
impossible to measure and inperceptible to us humans.
That's exactly why we make statement like you did. After a few
minutes of waiting, to us humans with a 70 year life span, it
appears as if nothing is happening (it's still sitting on that
same stinking move!) We think that after another few minutes
it is supposed to suddenly produce a brilliancy!
> In Go this is pretty common. Like when you play
> someone 9 stones weaker and they spend several minutes pondering over
> move and then come up something barely better than passing. If your
> eval does not know enough of situation more searching will not help.
If your program is scalable, it WILL help, but again you might have to
way a non-linear amount of time. It will only help in a practical
way if the program is within 3 or 4 stones of the required strength
to solve it quickly. Even in computer chess, a year of thinking
time cannot make up more than a few hundred ELO points.
Please note, it took several years for computers to go from 2000
strength to 2800 strength, and if you could use one of the 2800
strength
programs on that old 2 mgz 6502 that played 2000 strength, it would
take days or weeks of running per move to produce the same quality
moves.
I guarantee that if you WATCHED it make a move, you would be satisfied
after a few minutes that it wasn't going to play substantially better
but you would of course be wrong.
- Don
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
