D3QN-Driven Cooperative Partial Task Offloading for Latency-Energy Optimization in 6G LEO Satellite-Terrestrial Edge Networks

This work proposes CoPTO (Cooperative Partial Task Offloading), a three-segment LEO satellite-terrestrial edge computing framework that, unlike prior approaches which only choose between a single local device/edge server or an all-or-nothing satellite offload, enables a divisible computation task to be dynamically split for parallel execution between a serving terrestrial edge server (TES) and a neighboring TES — while retaining full satellite edge server (SES) offloading as an alternative when terrestrial conditions are unfavorable. This exploits previously unused spare capacity at neighboring terrestrial servers, which no existing DRL-based offloading scheme specifically targets.

A Dueling Double Deep Q-Network (D3QN)-based Intelligent Supervisor learns online to choose between full SES offloading and a continuous TES-TES split ratio, jointly minimizing latency and energy under per-task deadline and energy-budget constraints — a decision problem that is non-convex and time-varying, making it unsuitable for static optimization.

Simulation results show the approach reduces average normalized latency by up to 22.45% and energy consumption by up to 20.62% compared to a non-cooperative D3QN baseline, with the gains growing as task load increases (due to better mitigation of queue build-up at heavily loaded single servers).