Dependent-task directed acyclic graphs require each task to execute locally or offload to a mobile edge computing server, coupling latency, communication cost, and device energy. This study reports a controlled evaluation of deep reinforcement learning for offloading under a common simulator, dataset, and train/validation/final-test protocol. Baselines comprise a recurrent policy with proximal policy optimization (PPO-LSTM) and a double deep Q-network; extensions comprise a Transformer policy, discrete soft actor-critic, and stabilized PPO with advantage normalization, scheduled optimization, divergence-based early stopping, and orthogonal initialization. Agents and five heuristics are evaluated on final-test graphs of ten to fifty tasks under latency-oriented and energy-aware quality of experience (QoE) objectives, with five seeds and Holm-corrected tests. Stabilized PPO attained mean QoE of 0.31–0.43 across task-size and bandwidth analyses, though not at every size or bandwidth mean, exceeding Heterogeneous Earliest Finish Time (HEFT) by 0.05 to 0.15 absolute, yet its gain over PPO-LSTM remained about 0.001 and was not Holm-significant at any bandwidth. Both PPO variants reached the validation threshold at median update five, against ten for the Transformer. Ablation identified no single driver: only removing advantage normalization produced a Holm-significant drop, and only under the energy-aware objective. The Transformer remained competitive, whereas discrete soft actor-critic reduced energy rather than latency. Matching the recurrent baseline in size and inference time, stabilization remains viable on resource-constrained devices but cannot replace retraining on much slower links.
Copyrights © 2026