PDPPnet: Prioritized Delay-aware and Peer-to-Peer Task Offloading in Cloud-Edge Continuum with Double Dueling Deep Q-Networks

Efficient management of computational tasks in the cloud-edge continuum (CEC) is crucial for modern computing environments. This paper introduces the Prioritized Delay-aware and Peer-to-Peer Task Offloading (PDPPnet) scheme, employing multi-agent Double Dueling Deep Q-Networks (DDDQNs) to optimize task offloading in a distributed framework. PDPPnet uniquely addresses the challenges of uncertain load dynamics and task prioritization at edge nodes, enabling autonomous decision-making for non-divisible, delay-sensitive tasks without reliance on prior task models from other nodes. By formulating a multi-agent computation offloading problem, PDPPnet minimizes the expected long-term latency and task drop ratio while respecting task priorities. The architecture supports both peer-to-peer (P2P) and peer-to-Cloud (P2C) offloading, ensuring seamless task flow across the CEC. We integrated LSTM-predicted load dynamics with DDDQNs to enhance the long-term cost estimation, significantly improving decision-making efficacy. Simulation results demonstrate that PDPPnet markedly outperforms conventional offloading algorithms, reducing task drop ratios, average delay, and improving prioritized task throughput, thus optimizing the use of edge computational resources.