Bandit Learning-Based Distributed Computation in Fog Computing Networks: A Survey

Fog computing is a decentralized computing infrastructure that extends the capabilities of cloud computing closer to the edge of the network. In a fog computing network (FCN), computing resources, such as processing power, storage, and networking, are distributed at various points in the network, including edge devices, fog nodes (FNs) such as access points, gateways, and local servers. This architecture allows resource-limited end devices to offload part of their computational tasks to nearby FNs to achieve the reduced response delay services and energy efficiency. However, the high dynamics and complicated heterogeneity of fog computing environment in many application scenarios result in the uncertainty of network information that is raised as a critical challenge to design efficient computation offloading strategies. Meanwhile, existing solutions such as centralized optimization, matching and game theory-based decentralized offloading are inadequate to be adopted in this context because they require the perfect knowledge of system parameters. Considering as a promising approach to deal with the information uncertainty issues, bandit learning has used recently to develop distributed computation offloading (DCO) algorithms for the FCNs. In this paper, we aim at reviewing such of these state-of-the-art DCO solutions and elaborate their advantages and limitations. Additionally, we identify open research challenges and provide future directions for research in this area. Our survey shows that bandit learning is a promising approach for efficient computation offloading in fog computing, and we expect that future research will continue to explore its potential for improving the performance and energy efficiency of fog computing-enabled systems.