Yes, bad luck can be a problem.
Solutions:
1) RAVE/AMAF do bias good moves such that exploration take place anyway
2) Biased priors that initially forces many playouts for good
candidates, so that bad luck becomes less likely for moves that are
rated high using patterns or other means.
3) One can try to bias all moves to be searched initially if one has
no patterns
Valkyria uses 1 and 2. I used to have 3 but at some pointed I tested
it and it was really bad on large boards. Searching all moves at least
once (or more) on 19x19 wastes way too much for no gain.
But if the 1) and 2) does not work well because the program is weak
otherwise maybe 3 can be an option at least on small boards.
The hard part here is probable to have all these things working
simultaneously, and when it started to do so in Valkyria it was really
awesome! :-)
Nethertheless I some times observe some good moves not being searched
at all just because of random factors. I think there is a trade off
here. In order to get a really efficient search all of the time one
has to live with a small probability that some moves are overlooked
now and then.
Also highly selective search will correct itself given enough time,
because if the current best move is not good enough to win the winrate
will drop towards 0 which allows other move to be searched as well.
Magnus
Quoting Peter Drake <[email protected]>:
There has been some talk here of using a zero exploration coefficient.
Does this literally mean using the win ratio (with one "dummy" win per
node) to decide paths through the MC tree? It seems that the best move
could easily be eliminated by a couple of bad runs.
Does this only work when using RAVE/AMAF?
--
Magnus Persson
Berlin, Germany
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
