Reinforcement Learning based Matching for Decentralized Task Offloading in Fog Computing Networks

This paper proposes an algorithm called RL-MATCH for performing the distributed task offloading in the fog computing networks based on the matching theory and reinforcement learning. Fundamentally, RL-MATCH aims to match each task node (TN) having computation needs with each helper node (HN) with available computing resource for task offloading realization. Given the dynamic of IoT and fog computing environment featured by the time varying of task requirement and resource states, online learning is needed to estimate the network context as well as construct preference relations for the two sides of matching game. We, therefore propose to use Thompson Sampling (TS) method for this bandit learning problem to acquire better exploitation and exploration trade-off, therefore allowing TNs to achieve the informed preference relations of HNs quickly. Extensive simulation results demonstrate the potential advantages of the TS based learning over the ϵ-greedy and UCB based baselines.