A deep reinforcement learning based offloading scheme in ad-hoc mobile clouds

In this paper, we consider the problem of making an optimal offloading decision for a mobile user in an ad-hoc mobile cloud in which the mobile user can offload his computation tasks to nearby mobile cloudlets via a device-to-device (D2D) communication-enabled cellular network. We propose a deep reinforcement learning (DRL)-based offloading scheme which enables the user to make near-optimal offloading decisions by taking into account uncertainties of user's and cloudlets' movements and the cloudlets' resource availabilities. We first propose a Markov decision process (MDP)-based offloading problem formulation which considers the composite states of the user's and cloudlets' queue states and the distance states between the user and cloudlets as the system state space. The objective of the formulated MDP-based problem is to determine the optimal actions on how many tasks the user should process locally and how many tasks to offload to each cloudlet at each observed system state such that the user's utility obtained by task execution is maximized while minimizing the energy consumption, task processing delay, task loss probability and required payment. Then, we use a deep reinforcement learning scheme, called deep Q-network (DQN) to learn an efficient solution for the proposed MDP-based offloading problem. Extensive simulations were performed to evaluate the performance of the proposed offloading scheme. The simulation results validate the effectiveness of the offloading policies obtained by the proposed scheme.