Bayesian network transforms

[jpfairbanks]

I think we can do something really cool about converting deterministic code into Bayesian programs. From the data structures that we have we can put on every variable a distribution and then run the new program such that it creates and samples from distributions. We could probably do this with types like it will numbers where the part that would store the derivative instead of stores the randomness and the parts stores the value of the value. This would probably be a good use case for cassette.

[crherlihy]

We could also introduce the idea of uncertainty re: transition states, and/or the credibility of inputs (e.g. from robotic perception, for example).