[Computer-go] Bayeisan/Probablistic Playouts in Computer Go
alvaro.begue at gmail.com
Thu Sep 25 16:06:06 PDT 2014
I believe this has been discussed in the mailing list before: If your prior
distribution of the win rate of a move is uniform, after L losses and W
wins the posterior distribution will be a beta distribution with alpha=W+1
and beta=L+1. The expected value of this distribution is alpha/(alpha+beta)
= (W+1)/(W+L+2), which is equivalent to the common trick of starting the
counters W and L at 1 instead of at 0.
Of course one could start with a different prior, but I think staying
within the family of beta distributions makes sense because it's very
Is that the kind of thing you were looking for?
On Thu, Sep 25, 2014 at 6:28 PM, Alexander Terenin <aterenin at ucsc.edu>
> Hello everybody,
> I’m a PhD student in statistics at the University of California, Santa
> Cruz who previously worked on the Go program Orego, currently in the
> process of applying for the NSF fellowship. I am working on a Bayesian
> statistics - related research proposal that I would like to use in my
> application, and wanted to know if someone was aware of any research
> related to my topic that has been done.
> Currently, it seems most MCTS-based Go programs, in the playouts, treat
> the strength (win rate) of each move as a fixed, unknown value, which is
> then estimated using frequentist techniques (specifically, by playing a
> random game, and taking the estimate to be wins / total runs). Has anyone
> attempted to instead statistically estimate the strength of each move using
> Bayesian techniques, by defining a set of prior beliefs about the strength
> of a certain move, playing a random game, and then integrating the
> information gained from the random game together with the prior beliefs
> using Bayes' Rule? Equivalently, has anyone defined the strength of each
> move to be a random variable rather than a fixed and unknown value? Without
> making this email too long, there’s some theoretical advantages that might
> allow for more information to be extracted from each playout if this setup
> is used.
> If you are aware of any work in this direction that has been done, I would
> love to hear from you! I’ve been looking through a variety of papers, and
> have yet to find anything - it seems that any work remotely related to
> Bayes’ Rule has concerned the tree, not the playouts.
> Thank you in advance,
> Alex Terenin
> aterenin at ucsc.edu
> Computer-go mailing list
> Computer-go at dvandva.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go