TG-HWBO: Transformer-Guided Hybrid Whale-Bee Optimization and a Reproducible Benchmark Protocol for Semantic-Aware MEC Offloading
Mobile Edge Computing; Semantic Offloading; Transformer; Whale Optimization Algorithm; Artificial Bee Colony; MAPPO; DCEDRL; Multi-Objective DRL; QoE; 6G Edge Intelligence
AuthorsAbstractMobile Edge Computing (MEC) for novel 6G applications has to simultaneously optimize latency, energy consumption, semantic fidelity, security, and scalability while accounting for user-perceived quality. This article proposes TG-HWBO, a transformer-guided hybrid Whale Optimization Algorithm-Artificial Bee Colony approach to semantic-aware and risk-aware task offloading. A Temporal Convolutional Transformer (TCN-T) enables accurate predictions of task entropy, delay sensitivity, and threat-related priorities to steer WOA global search and ABC local exploitation. Semantic compression, encryption customization, gateway coordination, and QoE-aware prioritization are embedded into the hybrid optimization loop. The current article intentionally positions itself as a study of the architecture-and-benchmark-protocol design rather than a comprehensive state-of-the-art performance comparison. A preliminary evaluation of the implemented TG-HWBO algorithm against the five selected baseline methods is presented using the numerical results from the provided source study. In the reported experiments, TG-HWBO achieved 0.43-0.59 ms for communication latency, 1.9-3.9 J/task for energy consumption, 9.2-9.4 QoE, and 48.9-52.0% threat mitigation rate; with the task-drop ratio of 7.6% for 15,000 tasks. The findings are shared as preliminary results to establish the general trends, since the distribution across multiple runs and independent reimplementation’s are yet to be reported for all the methods. To support such comparisons, the article formally designs a common benchmark protocol for three recently proposed methods: semantic-aware multi-modal offloading (MAPPO-based), DCEDRL, and the multi-objective adaptive deep reinforcement-learning approach for 2025. The headline results from the original articles are not included in the current tables, since they do not directly correspond to the common set of metrics. The resulting manuscript provides a unified architecture description, a transparent preliminary evaluation, and sets the groundwork for a reproducible contemporary benchmark.
•••••••••••••••••••••••••••••••• ejprd.org - Published by Riset Publication Services LLC
EJPRD
Copyright ©2026 by Riset Publication Services LLC