Answers are grounded in this paper and cite their sources.
Task offloading in fog computing is difficult because nodes are heterogeneous, workloads are non-stationary, and failures are frequent. Autonomic computing offers self-management through the MAPE-K loop but relies on static rules, and heuristic schedulers degrade under non-stationary or adversarial workloads. Deep Reinforcement Learning (DRL) is promising, yet studies usually evaluate one algorithm in isolation without statistical validation. We introduce AFODRLS , a framework that realizes all four MAPE-K self-* properties in one loop driving four interchangeable DRL agents (DQN, DDQN, SAC, PPO), with an Adaptive Meta-Controller that selects between heuristic and DRL scheduling and an Online Capacity Estimator that replaces fixed capacity scaling. On a seven-level complexity benchmark (1,465 simulation trials), advanced heuristics outperform DRL at levels 1–4; at level 5 the best DRL agents match the best advanced heuristic; and at levels 6–7 the DRL group outperforms the heuristic groups, with DDQN (level 6) and DDQN/PPO (level 7) exceeding the best heuristic with large effect sizes (Cohen’s \(d>4.5\) , \(p<0.001\) ). The trained policies also transfer zero-shot to two real-world traces (GoCJ and HPC2N), matching the strongest heuristic on success rate without retraining.
No takes yet. Share an insight, caveat, or question.
Alturki et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: