
Research Article
Dynamic Offloading Based on Meta Deep Reinforcement Learning and Load Prediction in Smart Home Edge Computing
@INPROCEEDINGS{10.1007/978-3-031-54521-4_23, author={Mingchu Li and Shuai Li and Wanying Qi}, title={Dynamic Offloading Based on Meta Deep Reinforcement Learning and Load Prediction in Smart Home Edge Computing}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 19th EAI International Conference, CollaborateCom 2023, Corfu Island, Greece, October 4-6, 2023, Proceedings, Part I}, proceedings_a={COLLABORATECOM}, year={2024}, month={2}, keywords={Edge computing Task offloading Meta deep reinforcement learning Smart homes Dynamic trade-off}, doi={10.1007/978-3-031-54521-4_23} }
- Mingchu Li
Shuai Li
Wanying Qi
Year: 2024
Dynamic Offloading Based on Meta Deep Reinforcement Learning and Load Prediction in Smart Home Edge Computing
COLLABORATECOM
Springer
DOI: 10.1007/978-3-031-54521-4_23
Abstract
In the edge computing enabled smart home scenario. Various smart home devices generate a large number of computing tasks, and users can offload these tasks to servers or perform them locally. Offloading to the server will result in lower delay, but it will also require paying the corresponding offloading cost. Therefore, users need to consider the low delay and additional costs caused by offloading. Different users have different trade-offs between latency and offload costs at different times. If the trade-off is set as a fixed hyperparameter, it will give users a poor experience. In the case of dynamic trade-offs, the model may have difficulty adapting to arrive at an optimal offloading decision. By jointly optimizing the task delay and offloading cost, We model it as a long-term cost minimization problem under dynamic trade-off (DT-LCMP). To solve the problem, we propose an offloading algorithm based on multi-agent meta deep reinforcement learning and load prediction (MAMRL-L). Combined with the idea of meta-learning, the DDQN method is used to train the network. By training the sampling data in different environments, the agent can adapt to the dynamic environment quickly. In order to improve the performance of the model, LSTNet is used to predict the load level of the next slot server in real time. The simulation results show that our algorithm has higher performance than the existing algorithms and benchmark algorithms.