REMS: Resource-Efficient and Adaptive Model Selection in 5G NWDAF

Deep neural networks (DNNs) are widely used to meet the needs of service consumers. However, most research on DNN focuses on the use of single-task learning models. In this approach, each task necessitates distinct inference resources, which results in the allocation of separate computing resources. This increases the possibility of failing to meet the deadline for task completion due to insufficient computing resources, especially as the task request rate increases. To tackle this challenge, we propose a resource-efficient and adaptive model selection (REMS) scheme that adaptively selects a multi-task learning model (MTL) and single-task learning (STL) model. We formulate the model selection problem as a Markov decision process (MDP) to minimize resource consumption while satisfying latency requirements. The optimal policy can be obtained by converting the MDP problem into Q-learning. The evaluation results demonstrate that REMS achieves a significant performance improvement compared to the existing scheme.