Why not to implement .nextCategorical by resampling?

[garyzhubc]

.nextCategorical() is using inverCDF to simulate from categorical distributions. Was wondering why not to implement resampling here?

[alexandrebouchard]

Very good point, I had something like that in the back in my mind for some time. A simple solution could be a tree based method to store the probabilities. Alias method is another option but we would have to think if we need cheap update.