Abstract:We study the problem of predictive runtime monitoring of black-box dynamical systems with quantitative safety properties. The black-box setting stipulates that the exact semantics of the dynamical system and the controller are unknown, and that we are only able to observe the state of the controlled (aka, closed-loop) system at finitely many time points. We present a novel framework for predicting future states of the system based on the states observed in the past. The numbers of past states and of predicted future states are parameters provided by the user. Our method is based on a combination of Taylor's expansion and the backward difference operator for numerical differentiation. We also derive an upper bound on the prediction error under the assumption that the system dynamics and the controller are smooth. The predicted states are then used to predict safety violations ahead in time. Our experiments demonstrate practical applicability of our method for complex black-box systems, showing that it is computationally lightweight and yet significantly more accurate than the state-of-the-art predictive safety monitoring techniques.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the predictive runtime monitoring of black - box dynamic systems, especially the prediction in terms of quantitative safety properties. Specifically, the paper focuses on how to predict future states based on limited past - state observations and give early warnings of possible safety violations when the semantics of the system and the controller are unknown (i.e., in a black - box setting). This enables the system to take timely measures, such as switching to a backup safety controller, before danger occurs. ### Core Problems of the Paper 1. **Predictive Monitoring of Black - Box Systems**: Since the specific implementation details of the system and the controller are unknown, the system state can only be observed at a limited number of time points. 2. **Prediction of Quantitative Safety Properties**: It is not only necessary to predict future system states but also to evaluate whether these states will lead to safety violations. 3. **Real - Time and Accuracy**: The method needs to be lightweight enough to achieve real - time monitoring, and the prediction results should be as accurate as possible. ### Overview of the Solution The paper proposes a Taylor - based Predictive Monitoring (TPM) framework. This framework is implemented through the following steps: - **Learning Phase**: Use Taylor polynomials and backward difference methods to learn the approximate dynamic model of the system from past observed states. - **Predicting Phase**: Use the learned model to predict future system states and judge the future safety based on these predictions. ### Key Technical Points - **Taylor Polynomials**: Used to approximate system dynamics, with the formula: \[ P_l(s) = g(t) + \frac{g^{(1)}(t)}{1!}(s - t) + \frac{g^{(2)}(t)}{2!}(s - t)^2+\ldots+\frac{g^{(l)}(t)}{l!}(s - t)^l \] - **Backward Difference Method**: Used for numerical calculation of derivatives, with the formula: \[ \nabla^i x_0=\sum_{j = 0}^{i}(-1)^j\binom{i}{j}\frac{x_{-j}}{\tau^i} \] ### Experimental Verification The paper conducts experiments through two complex controlled dynamic systems (the F1Tenth racing environment and the F - 16 fighter environment) to verify the effectiveness and superiority of the TPM method, especially its superiority over the traditional Time - to - Collision (TTC) method in terms of prediction accuracy and computational efficiency. ### Summary The main contribution of the paper is to provide a lightweight and efficient predictive monitoring framework that can predict and warn of safety violations in advance in black - box dynamic systems, thereby ensuring the safety and reliability of the system.

Predictive Monitoring of Black-Box Dynamical Systems

Model Predictive Real-Time Monitoring of Linear Systems

Monitoring of Dynamic Process Using Hierarchical Probability Density Decomposition

Model Predictive Monitoring of Dynamical Systems for Signal Temporal Logic Specifications

System Safety Monitoring of Learned Components Using Temporal Metric Forecasting

Predictive Runtime Monitoring For Linear Stochastic Systems And Applications To Geofence Enforcement For Uavs

Kernel-based Learning for Safe Control of Discrete-Time Unknown Systems under Incomplete Observations

Learning-Based Approaches to Predictive Monitoring with Conformal Statistical Guarantees

Probabilistic Safety Constraints for Learned High Relative Degree System Dynamics

Safely Learning Dynamical Systems

Predictive Monitoring with Logic-Calibrated Uncertainty for Cyber-Physical Systems

Predictive Dynamical Systems

Event-Triggered Control for Safety-Critical Systems With Unknown Dynamics

Run-time Systems Failure Prediction Via Proactive Monitoring

Offline and online energy-efficient monitoring of scattered uncertain logs using a bounding model

Online Monitoring of Dynamic Systems for Signal Temporal Logic Specifications with Model Information

Reachability Analysis for Black-Box Dynamical Systems

Learning Run-time Safety Monitors for Machine Learning Components

Runtime Monitoring and Fault Detection for Neural Network-Controlled Systems