基于元强化学习的自适应卸载方法

An Adaptive Offloading Method Based on Meta Reinforcement Learning

计算卸载是移动边缘网络中的一个关键问题,基于深度学习的算法为高效生成卸载策略提供了一种解决方法。但考虑到移动终端设备的动态性以及不同任务场景之间的转换,需要大量的训练数据和较长的训练时间重新训练神经网络模型,即这些方法对新环境的适应能力较弱。针对这些不足,提出了一种基于元强化学习(Meta Reinforcement Learning,MRL)的自适应卸载方法,先对外部模型进行预训练,处理具体任务时再基于外部模型训练内部模型。该方法能快速适应具有少量梯度更新的样本的新环境。仿真实验表明,该算法能够适应新的任务场景,效果良好。

Computing offloading is a key problem in mobile edge networks.Deep learning-based algorithms provide a solution to efficiently generate offloading strategies.However,considering the dynamic characteristic of mobile terminal devices and the transformation between different task scenarios,a large amount of training data and a long training time are needed to retrain the neural network model,that is,these methods are weak in adapting to the new environment.In order to solve these problems,an adaptive offloading method based on meta reinforcement learning is proposed.Firstly,the outer model is pretrained,and then the inner model is trained based on the outer model when dealing with specific tasks.The method can quickly adapt to the new environment with a small number of gradient updated samples.Simulation results show that the algorithm can adapt to new task scenarios and has good effect.