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

Problem
The paper addresses a critical gap in the literature regarding the evaluation of safety in learned control policies, particularly in the context of quantum predictive control. Traditional safety filters can mask the true competency of the underlying policy by ensuring constraint satisfaction without assessing whether the policy itself has learned to operate safely. This work is particularly relevant as it is presented as a preprint, indicating that it has not yet undergone peer review.

Method
The authors introduce Intervention-Aware Variational Quantum Differentiable Predictive Control (IA-VQC-DPC), which employs a compact variational quantum circuit (VQC) policy. The training process incorporates a primal-dual intervention budget that penalizes excessive reliance on a differentiable Control-Barrier-Function (CBF) projection. The safety-attribution protocol developed in this work decomposes the corrections made during trajectory execution into contributions from the CBF and a runtime guard, allowing for a clear assessment of which components contribute to safety. The model is evaluated using high-fidelity BOPTEST building-control emulators, with experiments conducted over 5 seeds and 60 episodes per method.

Results
The results demonstrate that intervention-aware training significantly reduces the quantum policy’s raw pre-filter violations and total reliance on safety layers, both with p-values less than 10^-4. Notably, the quantum policy, constrained to approximately 400 parameters, outperforms a matched classical policy in terms of safety and comfort without incurring significant energy regression. The guard-off evaluation further substantiates that the improvements are attributable to the policy itself, revealing that a learned differentiable energy head is only safe when accompanied by a distribution-aware runtime guard.

Limitations
The authors acknowledge that their approach may be limited by the specific architecture of the variational quantum circuit and the assumptions inherent in the intervention budget. Additionally, the safety-attribution protocol, while generalizable, may require further validation across diverse applications beyond building control. The negative result regarding the learned energy head’s safety highlights a potential area for future research, as it suggests that safety cannot be guaranteed without appropriate runtime guards.

Why it matters
This work has significant implications for the field of safe reinforcement learning and quantum control, as it provides a framework for quantifying the contributions of learned policies versus safety mechanisms. The introduction of a measurable safety attribution protocol could influence future research directions in policy evaluation and safety assurance, particularly in complex environments where traditional methods may fall short. The findings underscore the importance of developing policies that inherently understand and manage safety, rather than relying solely on external filters. This is crucial for advancing the deployment of AI systems in safety-critical applications, as discussed in the context of quantum control in arXiv.