改进DQN的边缘计算任务卸载策略
Improved DQN edge computing task offload strategy
宋兴 1葛海波 1马世雄1
作者信息
- 1. 西安邮电大学电子工程学院,陕西西安 710121
- 折叠
摘要
为进一步提高边缘计算(MEC)中移动设备(MD)对低时延、低能耗计算卸载任务的需求,利用深度Q学习(DQN)、长短期记忆网络(LSTM)和注意力机制,设计一种基于DQN的深度强化学习卸载算法(LA-DQN).以最小系统总代价(时延和能耗加权和)为目标建立模型,使用一维残差卷积网络(Conv1D)和带有注意力机制的LSTM网络替换DQN网络的全连接层,提取MD的状态特征,减少需要计算的参数量并加强对输入状态的重点特征信息提取,加速算法收敛并得到最优卸载策略.仿真结果表明,与DQN、Full Local、Full Offload算法相比,LA-DQN算法能够有效降低任务处理的时延和能耗.
Abstract
To further improve the demand of mobile devices(MD)in edge computing(MEC)for low latency and low energy con-sumption computing offload tasks,a DQN based deep reinforcement learning offload algorithm(LA-DQN)was designed using DQN,LSTM and attention mechanism.A model with the goal of minimizing the total system cost(weighted sum of delay and energy consumption)was established,and one-dimensional convolutional residual network(Conv1D)and LSTM network with attention mechanism were used to replace the full connection layer of DQN network to extract the state characteristics of MD,reducing the amount of network parameters to be calculated and strengthening the extraction of key feature information of input state,the network convergence time was then accelerated and the optimal unloading strategy was obtained.Simulation results show that compared with DQN,Full Local and Full Offload algorithms,the LA-DQN can effectively reduce service processing delay and energy consumption.
关键词
边缘计算/深度强化学习/计算卸载/卸载策略/注意力机制/一维残差卷积网络/全连接层Key words
edge computing/intensive learning/computing unload/uninstall policy/attention mechanism/Conv1D/full connec-tion layer引用本文复制引用
基金项目
陕西省自然科学基金项目(2011JM8038)
陕西省重点产业创新链(群)基金项目(S2019-YF-ZDCXL-0098)
出版年
2024
NETL
NSTL
万方数据