SAPA: Structure-Adaptive Pattern-Guided Resource Allocation for Latent-Regime Edge-Intelligent Sensor Networks

Edge-intelligent sensor networks require limited sensing, computing, or communication resources to be allocated over many interacting sensing units under a fixed-sum constraint. The principal challenge is that the desired sensing pattern changes within an episode: early decisions emphasize high-risk nodes, middle decisions maintain connected coverage, and late decisions preserve reserve capacity, while the regime is unobserved. Existing score-ranking, ROI-based, and fixed-decoder policies often encode a single preference field and have limited capacity to switch sensing patterns under the nonnegative fixed-sum constraint. Structure-Adaptive Pattern-Guided Resource Allocation (SAPA) constructs focus, support, protection, and reserve patterns, then uses cue-conditioned gates and residual controls to mix them on the resource simplex. On a controlled 224-cell edge-sensing benchmark with 100k-step, three-seed training, SAPA obtains stable cumulative gain (13.23±0.04), compared with 6.97±0.05 for the strongest learned baseline, while also achieving the highest support capture among the compared methods.