A Multi-agent Reinforcement Learning based Offloading Strategy for Multi-access Edge Computing

As a front-end distributed computing paradigm, Multi-access edge computing (MEC) requires an efficient offloading strategy in the case of computation-intensive tasks. Traditional heuristic and reinforcement learning-based methods are limited by the number of edge serves. In this work, the task offloading in MEC is regarded as a multi-agent reinforcement learning (MARL) scenario, and an end-to-end model is developed to train the offloading strategy. The proposal is a decentralized framework, which is consistent with the edge computing servers. We fully consider the state sequences for edge servers and limited communication abilities, proposing a recurrent neural network as the communication module. Using a gating mechanism, we design a dual-recurrent network for combining the state sequences and historical communication results. Several experiments show that our method can achieve low latency and distributed data processing, and it outperforms heuristic methods and other reinforcement learning-based frameworks.