局部观测的车辆边缘计算在线服务迁移决策

Online service migration decision in vehicle edge computing under partial observation

针对多用户、全局信息缺失的复杂车辆边缘计算场景,提出了一种基于多智能体强化学习的车辆边缘计算服务迁移策略,使车辆在仅获得部分观察的情况下仍能进行不完整系统信息学习并进行分布式在线迁移决策。该策略使用Gumbel-Softmax采样改进的多智能体深度确定性策略梯度算法实现,用户间通过协作和竞争来达到共同的目标,从而使系统更加灵活,提高系统的整体收益和稳定性。真实数据集的仿真实验结果表明,该策略收敛速度快,在不同用户数量和任务到达率场景下均表现出较好的性能,鲁棒性和稳定性优于其他对比策略。

For complex vehicle edge computing scenarios with multiple users and global information scarcity,a vehicle edge computing service migration strategy based on multi-agent reinforcement learning was proposed.This allowed vehicles to learn incomplete system information and make distributed online migration decisions,even with only partial observations.This strateged employed a multi-agent deep deterministic policy gradient algorithm improved by Gumbel-Softmax sampling.Through collaboration and competition among users,a common goal was achieved,making the system more flexible and improving the overall benefit and stability of the system.Simulation results from real data sets show that this strategy converges quickly,performs well in scenarios with different numbers of users and task arrival rates,and exhibits superior robustness and stability compared to other comparative strategies.