I could find anything problematic with your specification so I just
make some comments.
Quoting Mark Boon <[EMAIL PROTECTED]>:
- When it reaches N simulations, the child of the root-node with the
best wins-visits ratio is played. I've also seen that simply the child
with the highest number of visits is selected. Is there a difference?
The following can happen if you chose the move with the best
wins-visits ratio: If the search is very selective (for example using
progressive widening or AMAF (RAVE)) a move can be searched a small
number of times and get a really good ration and selected. In Valkyria
which uses both methods of selectivity there is always moves that have
higher win-visits ratios than the most searched move (19x19).
If you do plain MC and UCT I think win-visits may be just as good as
highest number of visits, because the search will be quite wide.
Using the highest number of visits is sort of robust. Even if you know
the move is not as good as it initially seemed, you sort of know that
one has to search deep to see this and maybe the opponent cannot see
the refuting move. Also the reason for the low win-visits ratio may be
that problems in the position will give a low ratio for all moves if
they are searched deep enough. This is by definition true in any lost
position. If you can prove that all moves are losing the move that
required the largest tree for the proof is the candidate for provoking
a mistake.
Valkyria plays fast whenever win-visits and highest number of visits
agree and slow otherwise, in an attempt to eliminate uncertainty.
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?
Best
Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
