Actually, the example given (and the first element in the table) are
exactly what I stumbled upon. In the scope of MC, I think this style of
analysis is 100% correct for evaluation of leaf nodes. I guess time
will tell if assuming a conjugate distribution with prior
hyperparameters is a good approach. (I used terms from the link given)
Other variants of MC that look at the final score or similar
non-bernouli trials may benefit from the other distributions in the
table. I'm actually kind of surprised at the dissimilarity between the
normal and multinormal. I'd expect the multinormal to boil down to the
normal, but it looks like the standard normal has additional terms.
steve uurtamo wrote:
Maybe other simple solutions exist,
you might want to check out those distributions that magically
have nice properties with respect to the bayesian integral.
they're called conjugate priors, and lots of distributions have
nice, easy to calculate conjugate priors.
there's a table here:
http://en.wikipedia.org/wiki/Conjugate_prior
multivariate gaussians are very useful, and inverse wisharts
and dirichlets can be computed blindingly fast. this is
useful, for instance, in the case where you're trying to sample
directly from the parameter space of a hidden markov model
over multivariate gaussians, for instance, where you can
easily sample a huge number of trajectories to learn about the
parameter space.
sadly, the number of possible states in a go game is quite large.
s.
____________________________________________________________________________________
Food fight? Enjoy some healthy debate
in the Yahoo! Answers Food & Drink Q&A.
http://answers.yahoo.com/dir/?link=list&sid=396545367
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/