Value-based multi-agent deep reinforcement learning for collaborative computation offloading in internet of things networks

Han Li, Shunmei Meng, Jing Shang, Anqi Huang, Zhicheng Cai

Abstract: As a promising computing paradigm, mobile edge computing (MEC) can assist Internet of Things (IoT) devices in processing computation-intensive tasks. However, because of the different locations, limited resources, and dynamic loads of IoT devices and edge servers, designing a distributed computation offloading approach for the IoT system is challenging. In this paper, considering the delay-sensitive tasks and the binary offloading mechanism, we focus on designing a collaborative offloading scheme between different IoT devices. In order to reduce the overall consumed energy of IoT devices based on guaranteeing the delay constraints of tasks, we propose a distributed offloading algorithm by utilizing the value-based multi-agent deep reinforcement learning (MADRL) method. Specifically, with the individual networks deployed locally and the collaborative network deployed at the edge, IoT devices can learn their offloading policies from expericence data, and then determine their offloading decisions independently. Compared with several baseline algorithms, the proposed algorithm can better achieve our optimiaztion goal, which is demonstrated by the experiment results.