Profit Maximization of Delay-Sensitive, Differential Accuracy Inference Services in Mobile Edge Computing

作     者：Zhang, Yuncan Liang, Weifa Xu, Zichuan Jia, Xiaohua Yang, Yuanyuan 

作者机构：City Univ Hong Kong Dept Comp Sci Kowloon Tong Hong Kong Peoples R China Dalian Univ Technol Sch Software Dalian 116024 Peoples R China SUNY Stony Brook Dept Elect & Comp Engn Stony Brook NY 11794 USA 

出 版 物：《IEEE TRANSACTIONS ON MOBILE COMPUTING》 (IEEE Trans. Mob. Comput.)

年 卷 期：2025年第24卷第7期

页      面：6209-6224页

核心收录：

学科分类：0810[工学-信息与通信工程] 0808[工学-电气工程] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：University Grants Committee in Hong Kong (HK UGC) [7005845  8730103  9043510  9043668  9380137] 

主　　题：Edge computing differential-accuracy inferences approximation algorithms online algorithms model resolution instance placements multi-resolution service models primal-dual dynamic updating technique prediction mechanisms Edge computing differential-accuracy inferences approximation algorithms online algorithms model resolution instance placements multi-resolution service models primal-dual dynamic updating technique prediction mechanisms 

摘      要：The integration of Artificial Itelligence (AI) and edge computing has sparked significant interest in edge inference services. In this paper, we consider delay-sensitive, differential accuracy inference services in a Mobile Edge Computing (MEC) network while meeting user stringent delay and accuracy requirements. We formulate two novel profit maximization problems under static and dynamic settings of service request arrivals, with the aim of maximizing the accumulative profit of admitted requests. We assign differential accuracy service requests to the corresponding resolution instances of their requested service models, assuming that each resolution instance can serve up to L = 1 the same type of service requests. Since the profit maximization problem is NP-hard, we first formulate an Integer Linear Program (ILP) solution if the problem size is small or medium;otherwise, we devise a constant randomized algorithm with high probability. Then, we consider dynamic service request admissions without the knowledge of future request arrivals for a given finite time horizon, for which we develop a simple yet effective prediction mechanism to accurately predict the number of different resolution instances of each model needed, and pre-deploy the predicted number of resolution instances into cloudlets to reduce instantiating delays. We then devise an online algorithm with a provable competitive ratio for the dynamic profit maximization problem by leveraging the primal-dual dynamic updating technique. Finally, we evaluate the performance of the proposed algorithms by simulations. The simulation results demonstrate that the proposed algorithms are promising.