Reinforcement Learning based Matching for Parallel Computation Offloading in Dynamic Fog Computing Networks

Matching theory has been efficiently applied in fog computing networks (FCNs) to design distributed task offloading algorithms in the presence of selfishness and rationals of fog nodes. Given the dynamic nature of fog computing environment, it is challenging to obtain the stable matching since the preference relations of two sides of matching game is unknown a prior. To address this challenge, this paper proposes RL-MATCH, a framework for parallel computation offloading in dynamic fog computing networks (FCNs). RL-MATCH is based on the matching theory and Thompson Sampling (TS) empowered Multi-Armed Bandit (MAB) learning to deal with the inherent challenges allowing task nodes with needed computation tasks to estimate the informed preference relations of helper nodes with available computing resource quickly and accurately. Extensive simulation results demonstrate the potential advantages of the TS based learning over the
$\epsilon$
-greedy and upper confidence bound (UCB) based baselines.