Optimizing Resource Allocation for Joint AI Model Training and Task Inference in Edge Intelligence Systems

This letter considers an edge intelligence system where multiple end users (EUs) collaboratively train an artificial intelligence (AI) model under the coordination of an edge server (ES) and the ES in return assists the AI inference task computation of EUs. Aiming at minimizing the energy consumption and execution latency of the EUs, we jointly consider the model training and task inference processes to optimize the local CPU frequency and task splitting ratio (i.e., the portion of task executed at the ES) of each EU, and the system bandwidth allocation. In particular, each task splitting ratio is correlated to a binary decision that represents whether downloading the trained AI model for local task inference. The problem is formulated as a hard mixed integer non-linear programming (MINLP). To tackle the combinatorial binary decisions, we propose a decomposition-oriented method by leveraging the ADMM (alternating direction method of multipliers) technique, whereby the primal MINLP problem is decomposed into multiple parallel sub-problems that can be efficiently handled. The proposed method enjoys linear complexity with the network size and simulation results show that it achieves near-optimal performance (less than 3.18% optimality gap), which significantly outperforms the considered benchmark algorithms.