Probabilistic Shielding

Definition

Probabilistic shielding is a runtime safety-filtering method for RL that permits actions only when a probabilistic risk or cost budget can still be satisfied.

Why It Matters

Many safety constraints are not simple all-or-nothing forbidden-action rules. A policy may need to spend small amounts of risk in expectation while still keeping the probability of failure or cumulative expected cost under a threshold. Probabilistic shielding makes that budget explicit and enforces it during learning, not only after training.

Formalism / Key Objects

• In known-model unsafe-reachability shielding, the constraint is M_pi |= P_{\le p}(Reach(u)).
• The state risk beta_M(s) can be defined as the minimal probability of ever reaching an unsafe state from s:

$${\beta }_M(s)=\underset{\pi }{\min }P{r}_{M,\pi ,s}(Reach(u)).$$

• A sound shield uses an inductive upper bound beta with B_M(beta)(s) <= beta(s), so the shield never relies on an optimistic underestimate of risk.
• Risk-budget shields augment states from s to (s,q) or (s,x), where the second coordinate is the remaining probability or cost budget.
• In model-free CMDP shielding, a learned backup cost critic Q_b provides the risk estimate, and the shield-map blends the proposed policy distribution with a backup low-cost action when the distribution would overspend the current budget.
• Probabilistic Risk-Budget Shields records the formal known-model and model-free constructions.

Connections

• Extends Shielding from hard temporal-logic action filtering to budgeted stochastic safety.
• Depends on Probabilistic Model Checking when a transition model is available and the shield needs certified reachability bounds.
• Sound Value Iteration is a natural subroutine for computing upper bounds on unsafe reachability.
• Closely related to Constrained Markov Decision Processes, but probabilistic shielding enforces a constraint by transforming sampled actions rather than only penalizing violations in the objective.
• Related to Probabilistic Controlled Invariant Sets, which also packages probabilistic safety guarantees as a runtime safe action map.

Common Confusions

• A probabilistic shield is not merely a stochastic hard mask. It may permit risky actions when the remaining budget can absorb them.
• A shield based on a learned critic is only as reliable as the critic’s upper-bound quality; model-free shielding changes the safety assumption, not the need for a certificate.
• Preserving safety does not imply preserving the same exploration dynamics or reward-learning speed.

Key Sources

• HamelDeLeCourt2025 - Probabilistic Shielding for Safe Reinforcement Learning
• HamelDeLeCourt2025 - ProSh Probabilistic Shielding for Model-free Reinforcement Learning
• Quatmann2018 - Sound Value Iteration
• Alshiekh2018 - Safe Reinforcement Learning via Shielding
• Probabilistic Risk-Budget Shields