Generative Flow Networks for DAG-Aware Multi-Objective Task Scheduling in Edge–IoT Systems

Abstract

Efficient scheduling of dependent IoT tasks in Edge Computing is challenging due to heterogeneous resources, strict latency requirements, and inter-task precedence constraints. We model task dependencies as a Directed Acyclic Graph (DAG), which turns the scheduling problem into a combinatorial optimization challenge with an exponentially ample search space. To address this, we propose a Task-Dependency Aware Generative Flow Network (TD-GFlowNet) that incrementally constructs valid schedules while learning to sample them in proportion to a multi-objective utility function combining latency, energy consumption, and deadline penalties. Unlike traditional reinforcement learning methods that optimize a single trajectory at a time, TD-GFlowNet efficiently explores diverse scheduling solutions guided by the DAG structure. We benchmark our method against policy-gradient RL, Greedy, and Random baselines under physics-based edge settings. Experimental results show that TD-GFlowNet achieves superior latency-deadline trade-offs compared to RL, while being more robust to server heterogeneity than Greedy policies, thus establishing a generative paradigm for combinatorial optimization in edge task scheduling. Our contribution is not only the adoption of GFlowNets but their DAG-feasible, two-head factorization (task/server) with a physics-grounded multi-objective utility.