PMAF: A Framework for Static Analysis of Probabilistic Programs

PMAF is a framework for designing, implementing, and proving the correctness of static analyses for probabilistic programs. The framework, developed by Di Wang at Carnegie Mellon University, handles challenging features like recursion, unstructured control-flow, divergence, nondeterminism, and continuous distributions. It has been used to reformulate existing intraprocedural analyses and implement a new interprocedural linear expectation-invariant analysis, with experiments on benchmark programs demonstrating its practicality.

Use Cases

• Designing new static analyses for probabilistic programs based on the pre-Markov algebra framework.
• Implementing interprocedural analysis for Bayesian inference problems using the control-flow hyper-graph semantics.
• Creating procedure summaries for probabilistic programs with nondeterminism and continuous distributions.
• Evaluating the correctness and practicality of static analysis techniques on benchmark programs.

Strengths

• Framework supports analysis of complex program features including recursion, nondeterminism, and continuous distributions.
• Based on a novel semantics using control-flow hyper-graphs for each procedure.
• Has been used to implement three distinct analyses: for Bayesian-inference, Markov decision problems, and linear expectation-invariants.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.

Provenance

Source

Carnegie Mellon University

Collection Method

Likely contains benchmark programs and analysis results from the PMAF framework research.