Hmm step prediction

[hendriksanta]

Hello,
my first attempt at a pull request for tf probability - given a (pre-trained/ configured) HMM and an additional observation point, the function single_step_prediction outputs the probability over the hidden states (as a categorical distribution) and also returns a prediction for the next observed value (as a probs-weighted mean of the possible observation distributions).

In the initial step this is effectively the posterior_marginal (of the initial observation point). In a subsequent step, the function can then take the previous probability over hidden states (its output from the previous step) as a prior and a new observation point to update the probability over hidden states.
This way the function can be used to employ a trained HMM to make continuous predictions based on a stream of incoming (new) observation points.

Any suggestions/ tips as to what needs to be improved/ added very much appreciated.
Many thanks!