Markov decision processes with PyMC3

Questions
1

Hi there

I want to implement a reinforcement learning algorithm in Python and PyMC3 seems to be the most elaborate package for probabilistic modeling if have seem so far.
Has this been tried or are there any hints on how I could proceed? Am I in completely the wrong place ?

Thanks
Esteve

2

Have you checked out: https://github.com/sawcordwell/pymdptoolbox ? Maybe this may help?

3

I have. But the latest commit is from '15 and 225 GitHub stars vs. 3,369 is considerable :slight_smile:
I like the approach how probabilistic language is mapped onto python and I have the feeling MCMC and VI could be useful in this context too.

4

@ferrine has done some related research in the context of Bayesian RNN and LSTM. He said it is tricky and nothing works well.