Research map / Probabilistic and scientific modeling / Dynamical systems

Safety Certificate against Latent Variables with Partially Unidentifiable Dynamics

ICML 2025 · Probabilistic and scientific modeling › Dynamical systems · approach: Physics-informed dynamics learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes probabilistic safety certificates for confounded MDPs with latent variables, using a marginalized Q-function from causal RL (front-door adjustment) to enforce long-term safety probability with persistent feasibility, validated in a discrete driving simulation.

Problem

Ensuring long-term safety for stochastic systems with unobservable latent variables that cause partial unidentifiability and distribution shift between offline and online observed statistics, where existing safe control methods need full dynamics or perfect simulators.

Approach

Formulate invariance conditions in probability space via a marginalized value/Q-function defined on an auxiliary MDP with online statistics, estimated from offline data using front-door adjustment with an observable mediator; the resulting action constraint Q(y,u) - E_{pi}[Q(y,U)] >= 0 is always feasible and can be used by a myopic safety filter.

Key results

• In the discrete driving simulation (H=10, epsilon=0.2, 100 simulations), the proposed Q-function-based safety certificate achieves the safety objective, while DTCBF with offline statistics does not, even when its safety condition is satisfied.
• A toy example in the appendix shows the online safe probability for an action (0.55) can be much lower than the offline estimate (1.0), illustrating offline-statistics over-approximation.

Measured effects reported in the paper

Surprising findings

• Offline statistics can over-estimate safe probability of some actions, so standard offline-data-based barrier conditions can be misleading under latent-variable confounding.

Limitations

• Method requires discrete state, action, latent and mediator spaces and an observable mediator satisfying front-door assumptions.
• Evaluation is limited to a simplified discrete numerical simulation with no real-world or high-dimensional experiments.
• Proposed method assumes an unbiased Q-function estimate is available; the estimation error is not analyzed in the experiments.

Experimental setup

Datasets: simulated offline driving dataset generated from behavioral policy with latent road slipperiness

Benchmarks: simplified discrete driving simulation with latent road slipperiness

Metrics: probability of safety at each time step

Baselines: discrete-time control barrier function (DTCBF)

Builds on: discrete-time control barrier function front-door adjustment marginalized value function for confounded MDPs

Closest papers by meaning

• Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning (ICLR 2025 · similarity 0.85)
• Safely Learning Controlled Stochastic Dynamics (NeurIPS 2025 · similarity 0.84)
• Long-Term Safe Reinforcement Learning with Binary Feedback (AAAI 2024 · similarity 0.83)
• Probabilistic Shielding for Safe Reinforcement Learning (AAAI 2025 · similarity 0.83)
• Verified Safe Reinforcement Learning for Neural Network Dynamic Models (NeurIPS 2024 · similarity 0.83)
• A Provable Approach for End-to-End Safe Reinforcement Learning (NeurIPS 2025 · similarity 0.83)