Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Quantitative Predictive Monitoring and Control for Safe Human-Machine Interaction
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 a quantitative STL-U predictive monitor that outputs robustness degree intervals over Bayesian RNN flowpipes, and uses them for uncertainty calibration (a new loss) and adaptive control of a T1D insulin controller and a car-following controller.
Problem
Safe human-machine interaction requires predicting future states under human behavior uncertainty and monitoring safety requirements, but existing STL-U monitors only give qualitative satisfaction/violation, not the degree of satisfaction needed for fine-grained decisions.
Approach
Define quantitative STL-U semantics that compute a robustness degree interval (worst-case and best-case bounds) over all traces in a Bayesian RNN flowpipe, then use these intervals both to select SRT/dropout via a new loss and to adapt control actions.
Key results
- Proposed monitor raises overall average pre-alert time for adults from 1.2 to 23.0 minutes and F1 from 0.54 to 0.93.
- In T1D closed-loop simulation, time in range for adults rises from 90.9% to 96.3% (paired t-test p=0.01).
- Loss L_qt achieves F1 of 0.93/0.71/0.90 for adults/adolescents/children vs 0.88/0.60/0.81 for L_sat, but adolescent monitor F1 is slightly lower than baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use quantitative STL-U loss L_qt instead of L_acc | Bayesian LSTM uncertainty calibration · T1D requirement satisfaction, adults | F1: 0.66 -> 0.93 | Helped |
| Use quantitative STL-U loss L_qt instead of L_sat | Bayesian LSTM uncertainty calibration · T1D requirement satisfaction across adults, adolescents, and children | F1: 0.88 -> 0.93; 0.60 -> 0.71; 0.81 -> 0.90 | Helped |
| Use STL-U quantitative monitoring instead of monitoring the Bayesian LSTM mean trace | Predictive monitor · T1D hazards across adults, adolescents, and children | average pre-alert time: 1.2 -> 23.0; 9.2 -> 23.1; 10.6 -> 18.7 | Helped |
| Use STL-U quantitative monitoring instead of monitoring the Bayesian LSTM mean trace | Predictive monitor · T1D hazards, adults and children | F1: 0.54 -> 0.93; 0.88 -> 0.90 | Helped |
| Use STL-U quantitative monitoring instead of monitoring the Bayesian LSTM mean trace | Predictive monitor · T1D hazards, adolescents | F1: 0.78 -> 0.71 | Hurt |
| Add predictive monitoring and adaptive control to the Basal-Bolus Controller | T1D closed-loop management · Adults, adolescents, and children | time in range: 90.9% -> 96.3%; 80.9% -> 85.2%; 63.6% -> 74.8% | Helped |
| Add predictive monitoring and adaptive control to the Basal-Bolus Controller | T1D closed-loop management · Adults | number of hazards | Helped |
| Add predictive monitoring and adaptive control to the Basal-Bolus Controller | T1D closed-loop management · Adolescents and children | number of hazards | No clear effect |
| Add predictive monitoring and adaptive control to the Basal-Bolus Controller | T1D closed-loop management · All patient populations | hypo time: 4.8% -> 2.4%; 4.2% -> 3.3%; 29.4% -> 19.3% | Helped |
| Add predictive monitoring and adaptive control to the Basal-Bolus Controller | T1D closed-loop management · All patient populations | hyper time: 4.3% -> 1.3%; 14.9% -> 11.5%; 7.0% -> 5.9% | Helped |
Surprising findings
- The proposed predictive monitor has slightly lower F1 than the mean-trace STL baseline for the adolescent model (0.71 vs 0.78 overall).
- In the adolescent and child closed-loop T1D results, the reduction in hazards was not statistically significant (p=0.09 and p=0.33).
Limitations
- Adaptive controller is a proof-of-concept with hand-designed adaptation rules and thresholds.
- Only LSTM RNN models were evaluated; alternative RNN models are left to future work.
- Evaluation is entirely in simulation (UVA/PADOVA and CARLA), not real-world settings.
Experimental setup
Datasets: UVA/PADOVA T1D simulator data (30 virtual patients) CARLA car-following traces
Benchmarks: UVA/PADOVA T1D simulator closed-loop CARLA car-following closed-loop
Metrics: F1 score of requirement satisfaction average pre-alert time number of hazards time in range average distance average speed
Baselines: Basal-Bolus Controller L_acc loss L_sat loss STL monitor on mean trace
Builds on: STL-U signal temporal logic robustness Monte Carlo dropout Basal-Bolus Controller
Closest papers by meaning
- STL: Still Tricky Logic (for System Validation, Even When Showing Your Work) (NeurIPS 2024 · similarity 0.84)
- Safely Learning Controlled Stochastic Dynamics (NeurIPS 2025 · similarity 0.83)
- A Physics-Informed Machine Learning Framework for Safe and Optimal Control of Autonomous Systems (ICML 2025 · similarity 0.83)
- From Gambits to Assurances: Game-Theoretic Integration of Safety and Learning for Interactive Robotics (AAAI 2025 · similarity 0.82)
- SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning (NeurIPS 2025 · similarity 0.82)
- Learning Safe Control via On-the-Fly Bandit Exploration (ICML 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.