Safe Reinforcement Learning via Probabilistic Option Shielding
Summary
- Researchers propose Probabilistic Option Shielding (POS) to extend probabilistic logic shielding to the option level for reinforcement learning.
- POS recovers long-horizon safety without requiring a complete environment model.
- The method learns a model-free critic from one-step signals to estimate the safety probability of entire options under uncertainty.
- Evaluations in the LogiCity environment show POS achieves safer training while matching task performance of standard baselines.
Why it matters
POS advances safe reinforcement learning by combining long-horizon safety guarantees with model-free flexibility, improving training safety without sacrificing task performance.