Who Earns the Safety? Intervention-Aware Quantum Predictive Control with Safety Attribution

Researchers introduce Intervention-Aware Variational Quantum Differentiable Predictive Control (IA-VQC-DPC), a quantum machine learning framework that addresses a critical problem in safe reinforcement learning: distinguishing whether safety comes from the learned policy or from protective safety filters. The method uses Control-Barrier Functions with attribution protocols to measure true policy competence, demonstrating that quantum policies can achieve superior safety and comfort metrics compared to classical baselines at equivalent parameter budgets.

This research addresses a fundamental challenge in deploying learned controllers in safety-critical systems: the inability to determine whether a controlled system succeeds because the underlying policy learned proper behavior or because safety filters compensated for policy incompetence. Traditional approaches evaluate only post-filter outcomes, creating a measurement problem that obscures whether training actually improved policy quality. IA-VQC-DPC solves this through a primal-dual intervention budget that penalizes excessive reliance on Control-Barrier Function projections during training, forcing the quantum policy to develop genuine safety awareness rather than delegating responsibility to downstream filters. The safety-attribution protocol then decomposes corrections into filter-attributable and policy-attributable components, with guard-off evaluation stress-testing whether improvements persist without safety infrastructure. Testing on BOPTEST building-control emulators reveals the quantum approach reduces raw policy violations and filter reliance by statistically significant margins while maintaining energy efficiency. However, a critical negative result emerges: learned energy prediction models require runtime guards for safety, suggesting certain policy components cannot learn constraints reliably. This work impacts quantum machine learning development by establishing methodologies for honest performance evaluation in safety-constrained domains. The attribution framework applies beyond quantum systems, enabling better assessment of classical reinforcement learning approaches. For practitioners deploying learning-based controllers in building automation, power grids, or robotics, these methods provide tools to distinguish genuine policy improvements from filter-enabled failures, reducing deployment risks and improving transparency in autonomous system safety.

• →IA-VQC-DPC measures true policy safety by penalizing reliance on protective filters during training, preventing safety filters from masking incompetent learned policies.
• →Quantum policies achieve lower violation rates and reduced filter dependency than matched classical policies at equivalent parameter budgets on building-control tasks.
• →Safety-attribution protocol decomposes executed corrections to identify which system components earn safety credit, applicable across quantum and classical domains.
• →Guard-off evaluation reveals learned energy prediction heads require runtime guards for safety, establishing that not all policy components can learn constraints reliably.
• →The framework provides measurable methodology for safety-critical deployment of learning-based controllers in domains like building automation and autonomous systems.