Neural Combinatorial Optimization for Multiobjective Task Offloading in Mobile Edge Computing

• Published in: IEEE transactions on vehicular technology Vol. 74; no. 7; pp. 10869 - 10880
• Main Authors: Xiao, Xiang-Jie, Wang, Yong, Huang, Pei-Qiu, Wang, Kezhi
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.07.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Combinatorial analysis; Computation offloading; Computational modeling; Delays; Downlink; Edge computing; encoder-decoder model; Encoders-Decoders; Energy consumption; Finishing; Machine learning; Mobile computing; Mobile edge computing; Multi-access edge computing; multiobjective; Multiple objective analysis; neural combinatorial optimization; Optimization; Preferences; Real time; Real-time systems; Reinforcement learning; Servers; task offloading; Uplink
• ISSN: 0018-9545; 1939-9359
• DOI: 10.1109/TVT.2025.3546914

Task offloading is crucial in supporting resource-intensive applications in mobile edge computing. This paper explores multiobjective task offloading, aiming to minimize energy consumption and latency simultaneously. Although learning-based algorithms have been used to address this problem, they train a model based on one a priori preference to make the offloading decision. When the preference changes, the trained model may not perform well and needs to be retrained. To address this issue, we propose a neural combinatorial optimization method that combines an encoder-decoder model with reinforcement learning. The encoder captures task relationships, while the decoder, equipped with a preference-conditioned attention mechanism, determines offloading decisions for various preferences. Additionally, reinforcement learning is employed to train the encoder-decoder model. Since the proposed method can infer the offloading decision for each preference, it eliminates the need to retrain the model when the preference changes, thus improving real-time performance. Experimental studies demonstrate the effectiveness of the proposed method by comparison with three algorithms on instances of different scales.