A Distributed Computation Offloading Strategy for Edge Computing Based on Deep Reinforcement Learning

Mobile edge computing (MEC) has emerged as a new key technology to reduce time delay at the edge of wireless networks, which provides a new solution of distributed computing. But due to the heterogeneity and instability of wireless local area networks, how to obtain a generalized computing offloading strategy is still an unsolved problem. In this research, we deploy a real small-scale MEC system with one edge server and several smart mobile devices and propose a task offloading strategy for one subject device on optimizing time and energy consumption. We formulate the long-term offloading problem as an infinite Markov Decision Process (MDP). Then we use deep Q-learning algorithm to help the subject device to find its optimal offloading decision in the MDP model. Compared with a strategy with fixed parameters, our Q-learning agent shows better performance and higher robustness in a scenario with an unstable network condition.