Probabilistic Safety Verification of Neural Policies via Predicate Abstraction
Summary
Starter note seeded from metadata and title. This paper appears to verify safety properties of neural policies by constructing a predicate abstraction and reasoning about safety probabilistically.
Key Claims
- Full claims pending deeper ingest.
- Expected role in the wiki: bridge neural policy verification with abstraction-based formal analysis.
Methods / Formalism
- Predicate abstraction.
- Probabilistic safety verification.
- Neural policy analysis.
Evidence / Experiments
- Not yet reviewed in detail.
Connections
- Primary source for Predicate Abstraction in this vault.
- Important bridge from neural policies to Formal Methods.
- Likely useful for scaling safety beyond simple symbolic models.
- Cleanup triage (2026-05-22): high-priority full ingest because it links neural policies, Predicate Abstraction, and probabilistic verification. It is not superseded by shielding notes; the next pass should identify the verified property class, horizon assumptions, abstraction-refinement procedure, and how its probabilities relate to Probabilistic Model Checking.
Open Questions
- How are predicates chosen or refined?
- What is verified exactly: finite-horizon safety, trajectory-level properties, or something else?
Citation
Vinzent, M., Hermanns, H., and Hoffmann, J. (2026). Probabilistic Safety Verification of Neural Policies via Predicate Abstraction. AAAI-26.