Semi-symbolic inference for efficient streaming probabilistic programming

A streaming probabilistic program receives a stream of observations and produces distributions that are conditioned on these observations. Efficient inference is often possible in context using Rao-Blackwellized particle filters (RBPFs), which exactly solve problems when fall back sampling approximations necessary. While RBPFs can be implemented by hand to provide efficient inference, the goal programming automatically generate such implementations given input programs. In this work, we propose semi-symbolic technique for executing programs runtime system implements filtering. To perform exact approximate together, manipulates symbolic falls This approach enables implement same RBPF developer would write hand. ensure this, identify closed families – as linear-Gaussian finite discrete models guarantees inference. We have ProbZelus language. Despite an average 1.6× slowdown compared state art existing benchmarks, our evaluation shows speedups 3×-87× obtainable new set challenging benchmarks designed exploit families.