Gian-Carlo Pascutto wrote:
Don Dailey wrote:
Gian-Carlo Pascutto wrote:
If it is indeed a KGS flaw I may add a workaround to Leela as simple
as doing time = time / 10 as soon as winrate >95% or so. There is
still a possibility of losing on time then but it should happen less.
That is almost the identical heuristic that Lazarus uses. But in
addition to this the time is loaded up front - in other words early
moves use a lot more time - but when win-rate is high it's divided by
something not quite as severe as 1/10 since it gradually plays faster
and faster based on the move number. I think it's a very good
algorithm. Lazarus also would speed up when losing badly just
out of good manners - unless resignation is turned on.
I think I will add this and hope this resolves the problem.
But I still categorically object to the stance that it's the bots or
the programmers fault that it forfeits on time. As log as lag is not
compensated there is no way to avoid time losses, even if the bot
always moves instantly. You can at best improve the odds of this not
happening.
Agreed. CGOS tries to help by not counting the first 0.75 seconds as
reckoned by the server, but I don't pretend this is fair in any way nor
does it solve the problem. I do it mainly to make it possible for a
program to still complete a game even if it is virtually out of time
even in the face of small network delays. I think it's pretty common
for programs to lose 1/10 of a second or more on each move due to
various factors.
I don't object to a bot losing on time. In fact, I don't even object
to a bot taking extra time in lost positions. I think it's more civil
not to and I don't do that, but that's a bot authors choice. There
is no real reason NOT to take extra time when losing - you can reason
that since you are losing anyway, it doesn't hurt to take extra time to
possibly find a saving move. Of course it is definitely annoying on
CGOS to watch a program lose slowly and not resign or at least speed up
but it's within the programs right to do this.
- Don
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
