Hybrid Pluck: Exact Inference for Discrete-Continuous Probabilistic Programs via Mixed Knowledge Compilation
We present a new technique for exact inference in \emph{hybrid} probabilistic programs, those that make both discrete and continuous random choices. Our approach extends knowledge compilation, a popular technique for discrete inference, to support conjugate continuous latent variables and measure-zero observations. We formalize our approach for a first-order hybrid probabilistic language, and implement it as an extension to the Pluck programming language, called Hybrid Pluck. To evaluate our method, we assemble a new suite of 16 real-world hybrid inference benchmarks based on applied Bayesian models from a range of disciplines. On 14 problems for which existing exact systems time out after 10 minutes, Hybrid Pluck computes exact posteriors in milliseconds to seconds.