Thanks for the comments Magnus.
On 20-nov-08, at 13:00, Magnus Persson wrote:
I'd also like to hear opinions on what would be a good N for a
reference bot. If I set N to 2,000, just like the reference-bots on
CGOS, it plays rather poorly. Much worse than the ref-bot. I haven't
tested it a lot yet, but I think the break-even point against the
MC-AMAF ref bot would be somewhere around N=10,000.
What about doing a MC-AMAF-UCT version. Or perhaps just simply try
a MC-AMAF-TS in a best win-visits ratio first manner?
To start off, I wanted to keep things simple.
But to be honest, I hadn't given AMAF (or RAVE?) in combination with
tree-search much thought yet. Only yesterday did I look up an article
describing RAVE and it's actually not entirely clear to me yet how
this would be best implemented. I also did a little searching for
past posts to this mailing-list but it did little to clarify things
in my mind.
The way I understood the article, after a playout it updates all the
nodes at the current level of all the moves played during the playout
(if it's a win for the player) with a RAVE value that is used in a
similar fashion to the UCT value of a node. Only of the current node
does it update the win-visit ratio. Is that correct? This implies
creating a lot more nodes than I'm currently doing. I have seen
remarks that others postpone expanding a node until a certain number
of simulations have been done. I never quite understood the need for
that, but maybe this has to do with the fact that AMAF requires a
much larger number of node-creations?
The way I had implemented my UCT-search it became very intuitive how
to make the MC playout strategy prioritize certain moves in case of
heavy (or semi-light) playouts and needed zero modirfications to the
actual search algorithm. I'd probably need to completely review
things when a RAVE value starts to influence move-priority 'a priori'.
Mark
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
