A Multiagent Meta-Based Task Offloading Strategy for Mobile-Edge Computing

• Published in: IEEE transactions on cognitive and developmental systems 2024-02, Vol.16 (1), p.100-114
• Main Authors: Ding, Weichao, Luo, Fei, Gu, Chunhua, Dai, Zhiming, Lu, Haifeng
• Format: Article
• Language: English
• Subjects: Algorithm design and analysis; Algorithms; Computation offloading; Computational modeling; Data storage; Deep reinforcement learning; Deep reinforcement learning (DRL); Edge computing; edge task offloading; Energy consumption; Energy storage; Machine learning; meta-learning; Metalearning; Mobile computing; Multi-access edge computing; Multi-agent systems; multiagent; Multiagent systems; Network latency; Task scheduling
• ISSN: 2379-8920; 2379-8939
• DOI: 10.1109/TCDS.2023.3246107

Task offloading in mobile-edge computing (MEC) improves the efficacy of mobile devices (MDs) in terms of computing performance, data storage, and energy consumption by offloading computational tasks to edge servers. Efficient task offloading can leverage MEC technology to reduce task processing latency and energy consumption. By integrating the reasoning ability and machine intelligence of the cognitive computing architecture, such as SOAR and ACT-R, reinforcement learning (RL) algorithms have been applied to resolve the task offloading in MEC. To solve the problem that conventional deep RL (DRL) algorithms cannot adapt to dynamic environments, this article proposed a task offloading scheduling strategy which combined multiagent RL and meta-learning. In order to make the two actions of charging time and offloading strategy fully considered at the same time, we implemented a learning network of two agents on an MD. To efficiently train the policy network, we proposed a first-order approximation method based on the clipped surrogate objective. Finally, the experiments are designed with a variety of the number of subtasks, transmission rate, and edge server performance, and the results show that the MRL-based strategy has the overwhelming overall performance and can be quickly applied in various environments with good stability and generalization.