Multi-Stage Auction with Diffusion-Based DRL for Sequential-Tasks Offloading in Low-Altitude Economy Networks

In Low-Altitude Economy Networks (LAENets), drones operate as edge nodes that continuously acquire multi modal data for real-time decision-making, including road traffic, scene geometry, and dynamic elements, to support latency sensitive tasks such as large-scale traffic accident detection. How ever, these drone tasks are compute-intensive, often exceeding onboard capability, which necessitates offloading to ground Base Stations (BSs) with abundant communication and computation resources. Although existing studies have explored resource allocation and task scheduling in LAENets, most approaches consider drone tasks as monolithic, overlooking their inherent decompos ability. To overcome this issue, we propose a dependency-aware multi-stage offloading framework for LAENets, which decom poses large-scale drone tasks into a series of spatio-temporally related sub-tasks. Given drone mobility, we introduce a new sub tasks pre-migration model, allowing sub-tasks to be pre-migrated to the corresponding BSs before drones arrive at the target sub region to ensure the timeliness and continuity of overall task execution. To address the resource allocation and pricing problem of sub-tasks, we propose a novel interdependence-aware Multi-Stage Double Dutch Auction (MS-DDA) mechanism, which executes DDA auctions in each temporally coherent sub-region, ensuring that only winners of the previous auction stage can participate in subsequent auctions through cross-stage dependencies. Finally, to solve the bidding challenges of high-dimensionality, nonlinearity, and dynamic changes in MS-DDA, we employ a diffusion-based Deep Reinforcement Learning (DRL) algorithm that stabilizes exploration and accelerates convergence, achieving an average reward 26% higher than traditional SAC algorithms.