On Iteration in Discrete Probabilistic Programming
summary
Extends the discrete PPL dice with two concrete unbounded-iteration constructs and shows how exact inference still runs by compiling into binary decision diagrams and performing weighted model counting. One construct is a guarded while-loop that consumes a Boolean guard and is semantically serial on divergent runs; the other is ergodic fixed-point averaging, which computes the stationary distribution of the transition kernel and avoids that serial divergence. Includes a prototype integration into dice and empirical comparisons of convergence times and use cases.
AI-written summary · about
related talks
17mLazy Knowledge Compilation for Discrete PPLs
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
24mCompiling with Generating Functions
19mCompiling Probabilistic Programs for Variable Elimination with Information Flow
22mRoulette: A Language for Expressive, Exact, and Efficient Discrete Probabilistic(…)
15mScaling Optimization Over Uncertainty via Compilation
13mSemantic Foundations for Laziness in Discrete Probabilistic Programming
20mBit Blasting Probabilistic Programs