Xinhao Liandao
Radio-frequency neuromodulation has emerged as a promising therapeutic modality for treating neurological disorders,
offering precise spatial targeting and non-invasive delivery [1].
However, RF energy deposition in biological tissues raises
critical safety concerns, particularly regarding specific absorption rate (SAR) limits established by regulatory agencies [2].
Current clinical systems often employ conservative safety
margins that may limit therapeutic efficacy, motivating the
development of adaptive approaches that optimize the safety utility tradeoff.
RF dosimetry and safety analysis has been extensively
studied in the context of wireless communications [3] and
medical applications [4]. Traditional approaches rely on worst case analysis and conservative safety factors, often resulting
in suboptimal performance.
Constrained reinforcement learning has gained attention for
safety-critical applications [5], [6]. Primal-dual methods, in
particular, provide theoretical guarantees for constraint satisfaction while maintaining learning efficiency [7]. Recent work
has applied these techniques to robotics [8] and autonomous
systems [9], but applications to RF systems remain limited.
Beamforming optimization for medical applications has
focused primarily on unconstrained problems [10] or used
convex optimization with fixed constraints [11]. Our approach
bridges this gap by enabling adaptive constraint handling
through learning-based methods.