Might be nice to use a Dirichlet distribution to generate the parameters of the HMM randomly

[awillats]

When picking transition and emission probabilities, I generally hard-code these. That makes it harder to generate systems of arbitrary size for debugging. It also means we're less likely to see things like very low transition probabilities which can cause numerical issues for decoding.

Using a dirichlet distribution, we'd be able to set these parameters randomly every time.

https://en.wikipedia.org/wiki/Dirichlet_distribution

possible C++ implementation:
https://github.com/gcant/dirichlet-cpp

Downside: likely requires an additional library just for debugging

[awillats]

Could also be accomplished by generating matrix randomly, then normalizing rows to sum to 1. This is what Kevin Murphy's HMM toolbox does. That approach is less direct and less flexible, but simple to implement