Citation

Samad, Abdul and Sheraz, Muhammad and Chuah, Teong Chee and Lee, It Ee and Yeoh, Chun Yeow and Saeed, Sohail Imran (2026) TD3-based semantic communication for energy-efficient DNN inference offloading in MEC networks. Computer Networks, 290. p. 112851. ISSN 1389-1286

Text 1-s2.0-S1389128626008637-main.pdf - Published Version Restricted to Repository staff only Download (2MB)

Abstract

This work investigates computation-aware offloading for deep neural network (DNN) inference tasks in mobile edge computing (MEC) systems. To address limited wireless resources during task offloading, we adopt semantic communication (SC) in the form of DNN split inference, whereby users offload compact intermediate feature representations to the edge server for remote execution. Specifically, we characterize the relationship between the semantic compression ratio and the computation partition ratio for the adopted DNN architecture, enabling the offloading policy to capture the nonlinear tradeoff between local feature extraction and transmitted payload size. To achieve energy-efficient inference offloading, we formulate a joint optimization problem that minimizes total system energy consumption, including local computation energy, uplink transmission energy, and edge-server processing energy, by jointly optimizing local-edge computation partitioning and wireless bandwidth allocation. Since the resulting problem is highly coupled and non-convex, we design a reinforcement learning-based environment that captures the DNN inference structure, orthogonal frequency-division multiple access (OFDMA) resource allocation, and computation–communication constraints. The Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is then employed to learn continuous offloading and bandwidth-allocation policies through interaction with the simulated MEC environment. Simulation results show that the proposed SC-enabled computation-aware offloading framework achieves lower total system energy consumption than representative deep reinforcement learning baselines and a non-learning greedy heuristic, particularly under bandwidth-constrained wireless conditions.

Item Type:	Article
Uncontrolled Keywords:	Mobile edge computing, DNN inference, Computation offloading, Deep reinforcement learning, TD3, Energy minimization
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5101-6720 Telecommunication. Including telegraphy, telephone, radio, radar, television
Divisions:	Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	30 Oct 2026 02:39
Last Modified:	30 Oct 2026 02:39
URII:	http://shdl.mmu.edu.my/id/eprint/17029

Downloads

Edit (login required)