So, you use AMAF for "simulating" the first UCT evaluations ?
I though the classical way to use AMAF, was to affect only the
win/lose ratio portion of the uct equation.
Obvioulsy it should be allowed to use an arbitrary
large number of AMAF simulation accumulating them longer
than what it take to expand a node.
I think that a classical way to affect the win/ratio is to
decrease the effect of the AMAF correction as the number
of simulation grows.
If you test with a very low number of simulation
(in the 1000 - 3000 range), i think you should be
able to get out a very nice improvement out of the
AMAF version. If you don't, i would think that something
is wrong somewhere.
What test process do you use for this version ?
> From: [EMAIL PROTECTED]
> Date: Thu, 27 Nov 2008 18:15:34 -0200
> To: computer-go@computer-go.org
> CC:
> Subject: [computer-go] Monte-Carlo Tree Search reference bot
>
> I have added a MCTS implementation to my reference-bot project. It
> allows you to specify how many nodes to search, after how many nodes
> to expand and whether to use AMAF or not. If you specify the number
> of nodes before expansion to be the same as the total number of nodes
> to search and set AMAF to true you get the equivalent of Don's
> original ref-bot. When you set AMAF to false and set the number of
> nodes before epxansion to 1 you get my original UCT search algorithm.
>
> Between those extremes there might be a good formula to combine AMAF
> with tree-search, but unfortunately I haven't had time lately to look
> for one. The few small attempts I made show no benefit using AMAF in
> tree-search, only when used on a single level. The contrast between
> the two exptremes is very stark, so I'm actually convinced there
> should be a way to use both. This implementation is also quite a bit
> slower than my original search algorithm but I also didn't have time
> yet to trace it. It might simply be due to the different expansion
> method, which is much more expensive with a value of 1. Also,
> searching for the best UCT node gets more expensive with more
> (unused) nodes on each level. Using a higher expansion value may
> alleviate the performance hit. Anyway I think this is a reasonable
> starting point.
>
> At first I intended to create a different project for the search
> reference bot. But half of the code (the MC part) is the same. And I
> don't want to end up having to maintain the same code in two places.
> I also didn't want to separate out some code into a separate library
> and making the structure for the simple ref-bot more complicated.
> This organization may need some more thought though.
>
> I'll update the project pages tomorrow.
>
> Mark
>
>
>
> _______________________________________________
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
_________________________________________________________________
Email envoyé avec Windows Live Hotmail. Dites adieux aux spam et virus, passez
à Hotmail ! C'est gratuit !
http://www.windowslive.fr/hotmail/default.asp
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
