I am not entirely sure what you mean here by tuning coefficients do the
heuristics in question require some form of parameterization? How are
these parameters tuned?
There are patterns in the tree and in the Monte-Carlo simulations. Also,
some other expertise (not encoded as pattern) can be used. Mixing all
these patterns/tricks imply coefficient (typically, one coefficient
quantifying the bias induced by a given pattern).
Then, for each new coefficient, one can play plenty of games with each
value of the coefficient, and pick up the best value. In particular for
the MC,
we have nothing better than that. A puzzling trouble is that the optimal
value strongly depend on the computation time you have per move, the
size of the board, and the configuration (number of cores, number of
nodes).
Also, I'm afraid that coefficients can't be picked up in one
implementation and put in another implementation, unless the
implementations are very related - so many small details have
such a strong influence.
This seems to be the case and I still do not really on some level
understand why. Since with the chinese go rules the board should be
effectively stateless (exempting ko information) all the information be
contained in the the current position. Why additional information is
needed i.e. an annotation of the last played move is required on some
level is a mystery to me.
One can build nice examples of that for some artificial games: the
knowledge of the last move helps *for finite computational cost*.
Sylvain has shown interesting elements around that, but as far as I know
this was not in his ph.D. thesis and this is not published.
I am sure I understand the distinction here between patterns in the
tree and patterns used in the heavy playouts. I guess by this you mean
they are not the same thing.
Yes, for sure. The bias induced by a pattern (whenever it is the same
pattern) is very different in the tree and in the MC part.
Olivier
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
