On 2/9/07, Weston Markham <[EMAIL PROTECTED]> wrote:
I don't seem to have any numbers on this anymore, but I should be able
to try some experiments this weekend. I do have some code that does
what I describe below. It is also using an "all moves as first"
heuristic. According to my notes, I made this change in an attempt to
avoid severely conservative (in my non-expert opinion) moves near the
beginning of the game, which seem to be preferred when using
all-moves-as-first. It specifically aims for a 30-point lead at the
beginning of the game, and reduces this by one point for each turn
into the game.
I should point out that I am not averaging scores, but simply changing
which games I count as wins for the evaluation of a move. This is
perhaps not quite what Steve Uurtamo had in mind when he was
originally musing about being greedy at the beginning of the game.
Nevertheless, it is a very similar sort of idea to what he described,
so I thought that I would mention it.
Weston
On 2/8/07, Chris Fant <[EMAIL PROTECTED]> wrote:
> On 2/8/07, Weston Markham <[EMAIL PROTECTED]> wrote:
> > I believe that I have had some success with an approach like this,
> > actually. I believe that I initially only tally games that are won by
> > a certain margin, and reduce that margin to zero as the game
> > progresses. I am pretty sure that this improves upon straight Monte
> > Carlo. I think that I can get some numbers on it, if anyone is
> > interested.
> >
> > Weston
>
> Yes, please.
I did run some tests over the weekend, but they did not complete as
quickly as I had expected. Anyway, I have enough data at this point
to say that although there is nothing spectacular to show one way or
the other, I was probably wrong that the code was an improvement.
Playing against gnugo 3.7.10, komi 7.5, playing black in 50% of the games:
86/732 games won baseline (11.7%)
106/1000 games won after enabling logic described above. (10.6%, a
slight degradation)
I also tried a version that dynamically adjusts an "expected score",
and tallies the number of games where the score exceeds this. It
appears to give an improvement: 12.8% of the games were won, out of
1000.
Weston
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
