I don't think this specific test had been done. But I'm assuming the result will be the same as previous tests: deviating from the pursuit of the the highest
winning percentage leads to a degradation in strength.
Brett Koonce wrote:
Greetings from a lurker,
Forgive me if I am talking out of
Part of the problems stem from that playouts are weak, and more
specifically notably weaker than the program itself.
To begin with, a consequence is that most areas of the board are less
clear than they should to playouts. This entails, I think, a preference
for probable points against sure
MC is playing most goal-directed (zielgerichtet
in German) when the position is balanced or when
the side of MC is slightly behind. However, when
MC is clearly ahead or clearly behind it is playing rather
lazy.
At some point we were investigating that here, but only on small sets of
games
In the last few weeks I have been investigating
Monte-Carlo (MC) game tree search on a rather abstract
level. Especially I was able to reproduce the
following behaviour of MC in a very clear model:
MC is playing most goal-directed (zielgerichtet
in German) when the position is balanced or when
Especially I was able to reproduce the
following behaviour of MC in a very clear model:
MC is playing most goal-directed (zielgerichtet
in German) when the position is balanced or when
the side of MC is slightly behind. However, when
MC is clearly ahead or clearly behind it is playing
I have some sense that it might be possible to slightly improve the
playing strength with some dynamic komi scheme. However, I have also
experimented quite a bit with various ways to do this and in each case I
have been able to detect at least a slight weakening of play.
That probably just
Don Dailey wrote:
That probably just means I have not stumbled on the right ideas or that
I was not able to properly tune it. I would be delighted if someone
was able to show us a workable scheme. I believe if something is found
it will result in a very minor improvement, but that it will
Hello Gian-Carlo,
Gian-Carlo Pascutto wrote:
There has been discussion here about dynamic komi to keep
the winning rate close to 50%. As far as I saw there was no
clear conclusion about whether that works. Some people argued
that it should not exist and measuring objective winning rates is
On Mon, 2008-09-08 at 20:01 +0200, Gian-Carlo Pascutto wrote:
Don Dailey wrote:
That probably just means I have not stumbled on the right ideas or that
I was not able to properly tune it. I would be delighted if someone
was able to show us a workable scheme. I believe if something is
Interesting analysis, Don.
Human players sometimes adhere to a simple policy: rich men don't pick fights.
When one is objectively far ahead, one picks up the easy profits, and otherwise
takes no risks. If moves A, B, and C are comparable risk-wise, one would prefer
the more profitable of the
Don Dailey wrote:
Would a discrepancy on the amount of ELO gained or lost per handicap
stone, when comparing MC bots to humans classical computers, be a good
measure of the maximum possible improvement?
Maybe. How could you accurately make such a measurement without
thousands of games?
Actually your summary of what people do sounds exactly like what MC
programs do, except for one point...
MC programs don't differentiate moves by point value. They only look
at winning rate. It's extremely tough to differentiate the one move
sequence with 99.1% win rate when all other
Greetings from a lurker,
Forgive me if I am talking out of my hat. It has been a long time
since I have done any real coding.
It seems most of the gains in MC/UCT come fairly quickly (or rather
you can get within 50% of a good move guess with a few iterations).
It would be interesting
13 matches
Mail list logo
