Belief space binning

[WhiffleFish]

We cluster beliefs according to suitable features and use previously computed values of beliefs in the same cluster as b to predict the value of b. This allows us to learn which parts of the belief space is worth exploring. Currently, we use the initial upper bound and the entropy of b as the features and discretize the belief space into a finite number of bins according to these two features. The average value of beliefs in a bin is used as the prediction for the value of any new belief falling into the bin. If a bin is empty, the initial upper bound of the new belief is used as its predicted value.
-- SARSOP Paper

C++ Implementation