We present a new algorithm for Bayesian inference over probabilistic programs, based on data flow analysis techniques from the program analysis community. Unlike existing techniques for Bayesian inference on probabilistic programs, our data flow analysis algorithm is able to perform inference directly on probabilistic programs with loops. Even for loop-free programs, we show that data flow analysis offers better precision and better performance benefits over existing techniques. We also describe heuristics that are crucial for our inference to scale, and present an empirical evaluation of our algorithm over a range of benchmarks.

PDF file

Details