Starvation-Aware Priority Offloading and Resource Orchestration for Heterogeneous Time-Sensitive Services in Wireless Edge Networks

Wireless edge networks increasingly support heterogeneous time-sensitive services whose terminals compete for limited communication and computing resources, making it challenging to balance system-wide efficiency, priority-aware service guarantees, and starvation mitigation. However, many existing offloading studies mainly emphasize average performance and provide limited explicit modeling of task priority and long-term service insufficiency, which may weaken differentiated service provisioning and starve low-priority tasks under strong priority bias. To address these issues, this paper proposes PA-MAPPO, a starvation-aware multi-agent partial offloading framework under centralized training and decentralized execution. PA-MAPPO jointly optimizes task offloading ratios and transmission power, embeds task priority into environment-side scheduling, local observation encoding, and hierarchical reward shaping, and introduces a priority-age-based anti-starvation mechanism: a bounded, exponentially discounted class-level service-deficit state that activates virtual-priority compensation when a priority class remains persistently under-served. Simulation results show that PA-MAPPO attains the lowest unweighted processing cost among the learning-based methods that sustain differentiated service (2.59 vs.2.68 for the scheduler-only MAPPO reference), preserves the monotonic priority ordering across the evaluated training seeds, and guarantees compensation activation within a closed-form latency bound. An attribution study further separates the contributions of the two layers: the environment-side scheduler establishes the service hierarchy, whereas the complete priority-aware learning stack delivers the resource-efficiency gains under it.