Learning to optimize computation offloading performance in multi-access wireless networks

In this paper, we investigate computation offloading in a multi-access wireless network, which supports both cellular and WiFi connectivity between a mobile user (MU) and the edge server. The MU decides to process an arrived computation task locally at the device or offload it to the edge server for remote execution. The technical challenges of designing a computation offloading policy lie in the network uncertainties due to the MU mobility, the sporadic task arrivals, the spatially distributed WiFi connectivity and the intermittent wireless charging opportunities. Accordingly, we apply a Markov decision process framework to formulate the problem of computation offloading over the infinite discrete time horizon. The objective of the MU is to find a policy to minimize the expected long-term cost. Without the knowledge of network uncertainty statistics, this paper makes the first attempt to exploit the model-free DQNReg, which is built upon a deep Q-network by adding a weighted Q-value to the squared Bellman error, to solve an optimal computation offloading policy. Experiments validate the superior performance from our approach compared to the baselines in terms of average computation offloading cost.