Joint Service Placement and Resource Allocation for Long-Term DNN Inference Accuracy in Dynamic MEC Networks

By leveraging deep neural networks (DNNs), mobile edge computing networks can integrate advanced intelligent computing capabilities. Due to temporally dynamic changes in wireless channel and service requesting distribution, the long-term inference accuracy of DNN services can be severely affected over time. To address the above challenge, this correspondence aims to maximize the long-term inference accuracy by jointly optimizing service placement, bandwidth allocation, and wireless device association. By modeling the relationship between data size and inference accuracy, we employ regression techniques to derive the fitting curve for the deployed services. A two-timescale long-term optimization problem is transformed into a series of subproblems using Lyapunov analysis. We propose an alternating optimization algorithm to tackle with the subproblems, in which the convex-concave procedure is utilized for bandwidth allocation, while the branch-and-bound method is employed for service placement. Moreover, a low-complexity penalty-based method is further developed for service placement. Simulation results show that the proposed methods outperform baselines in inference accuracy and system fairness.