Learning-Based Computation Offloading for Edge Networks with Heterogeneous Resources

Mobile edge computing (MEC) has shown its potential in serving computation intensive tasks via offloading. However, the heterogeneity of MEC systems and the dynamic nature of wireless environment pose a great challenge to the design of offloading policies. In this paper, we investigate this computation offloading problem, where the heterogeneities of computational resource, channel state, task type and input data size are considered. We first propose a greedy algorithm, in which each arrival task is greedily offloaded to the edge server with minimal utility, based on a global information of network states. While this greedy algorithm performs well in terms of system utility, the overhead incurred to collect the global information is large, especially in dense MEC scenarios and time-varying channel scenarios. Inspired by this observation, we then propose a model-free offloading algorithm based on reinforcement learning, which does not rely on such kind of information and can make offloading decisions based on learning experience. By so doing, the communication overhead can be largely reduced. Extensive simulations show that the two proposed algorithms have similar performance in terms of system utility and can decrease the system utility by up to 50% compared with two widely used algorithms. The robustness of the two proposed algorithms is further verified.