基于分布式深度学习的多星计算卸载策略

在边缘计算增强的低轨卫星网络场景下，低轨卫星集群协同处理地面任务能有效降低用户响应时延。对卫星集群的联合卸载决策和资源分配优化问题进行研究，将其描述为一个混合整数规划问题，并采用了一种基于分布式深度学习算法的卫星边缘计算卸载算法（deep learning based offloading algorithm，DLOA）。该算法使用多个并行DNN用于生成卸载决策并采用经验回放存储新生成的卸载决策，当采用隐藏层结构不同的DNN，收敛速度比同构DNN提升18%，收敛值与最优值的比值基本为1，可以认为已收敛至最优。此外，探讨了DNN的数量对所使用的算法的影响，仿真结果表明采用少量DNN就可以获得近优的收敛效果。通过对不同任务规模下采用不同算法的任务完成率进行研究，结果表明DLOA算法可通过采用异构DNN和优化资源分配方案显著提升完成率，其较单星运算方案任务完成率提升1倍，较二进制粒子群算法方案提升20%。

Multi-satellite task offloading method based on distributed deep learning

In edge computing enhanced LEO satellite networks，collaboratively processing ground tasks by satellite clusters could effectively shorten the user response delay.The optimization of joint offloading decision and resource allocation of satellite clusters was studied，and was described as a mixed integer programming problem.A satellite edge computing enhanced deep learning-based offloading algorithm（DLOA）was adopted.The algorithm used multiple parallel DNN networks to generate offloading decisions and adopted experience replay to store the newly generated decisions.Particularly，the convergence rate of heterogeneous DNN with different hidden layer structure is 18% higher than that of homogeneous DNN with same hidden layer structure，and the ratio of convergent value to optimal value is basically 1，which can be considered to have converged to the optimal.In addition，the influence of the number of DNN on the proposed algorithm was discussed，and the simulation results show that the near-optimal effect can be obtained by using a small number of DNN.By studying the task completion rate of different algorithms under different task scales，the results show that DLOA algorithm can significantly improve the completion rate by using heterogeneous DNNs and optimized resource allocation schemes，which is twice as high as the single satellite scheme，and 20% higher than the binary particle swarm algorithm.