Hi Aja,
that's a good question. At least for the LGR policy without forgetting
(https://webdisk.lclark.edu/drake/publications/drake-icga-2009.pdf),
only using the first appearance of a reply did not significantly differ
in performance. A possible explanation could be that in cases where the
same move by the same player appears twice in a playout, the first stone
must have been captured, and therefore the answer to the second play is
the one that really influences the final position/result. I'm not sure I
repeated this experiment with LGRF, but I did try dismissing the tails
of playouts (with the rationale that there might be too much noise) and
ignoring stones that would later be captured (with the rationale that
those moves might be bad on average). Both variants were significantly
weaker than plain LGRF.
It's only a few lines of code, test it and see if it makes a difference
for your playout policy and program architecture. Stronger playout
policies than Orego's will have different interactions with LGRF. You
could even try saving several sets of replies per intersection, for the
first, second, third appearance of the previous move in a playout, in
the hope of capturing certain tactical situations with sacrifices. But I
don't expect much.
Hendrik
Am 26.01.2011 14:13, schrieb Aja:
Hi Hendrik,
Thanks.
Congratulations, you have done a really nice work. I check your
thesis. My result is consistent with yours of LBR-2. No benefit at
all, so I took it off. I adapt LGR-1 to softmax policy of Erica.
Basically, I am tuning the probability offset by checking some
aritifical test-positions. In 3000 playouts, now it scores around 57%
after 500 games, almost 60%, which is my target (my intuition is LGR-1
should help a lot already). :)
Actually I have one question and still can't figure out your
reasoning. In a playout, why do you over-write the earlier replies by
the later ones? Using the earliest one looks more reasonable to me.
Aja
----- Original Message ----- From: "Hendrik Baier" <>
To: <[email protected]>
Sent: Wednesday, January 26, 2011 5:00 PM
Subject: Re: [Computer-go] Computer-go Digest, Vol 12, Issue 79
Hi Aja,
I would be interested in your results. I think the LGRF policy is
only a small first step into the direction of more adaptive playouts
(and hopefully the overcoming of the horizon effect).
As for the Last-Bad-Reply idea, you can read about my experiences
with this and related policies in my Master's thesis, if you're
interested. It contains the idea that resulted in the "Power of
Forgetting" paper as well.
http://www.ke.tu-darmstadt.de/lehre/arbeiten/master/2010/Baier_Hendrik.pdf
regards,
Hendrik
I admit that it's difficult for me to include such deterministic
default policy. :-)
With softmax policy, using the information of "last-LOST-reply" is
maybe a good direction.
Aja
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
