DAGMARL-Fog: Graph-Aware Multi-Agent Learning for Adaptive and Sustainable Fog Task Scheduling

Latency-dependent IoT systems can be supported by fog computing; however, problems such as high control overhead, late synchronization, and non-scalability arise from centralized scheduling approaches. This paper introduces DAGMARL-Fog, a decentralized framework of multi-agent reinforcement learning with graph-aware state representation for scheduling tasks within IoT, fog, and cloud environments. The proposed framework uses graph representation of states, feasibility-filtered MARL, consensus-based neighborhood choice, and objective weight adjustment to improve latency, energy, load balance, reliability, and carbon emission. Simulations based on iFogSim2, latency modeling through EdgeCloudSim, and Google Cluster Trace achieve 132 ms latency, 7900 J energy, 95.6% deadline fulfillment rate, and 0.94 fairness index.