Deep Reinforcement Learning-Based DAG Task Scheduling for Heterogeneous Edge–Cloud Systems

Task scheduling for directed acyclic graph (DAG) applications in heterogeneous edge–cloud systems is challenging due to coupled computation–communication overhead and complex precedence constraints. This paper studies the makespan minimization problem and proposes DRL-ECTS, a deep reinforcement learning–based scheduling algorithm that learns server assignment policies for DAG tasks. The scheduling process is formulated as a Markov decision process, where the agent observes task/graph features and server states and selects a target server for each ready task. To improve exploration and training stability, we adopt an entropy-regularized policy optimization scheme and further introduce a behavior cloning mechanism in the early stage using demonstration trajectories, which gradually decays to autonomous reinforcement learning. We evaluate DRL-ECTS on a simulated edge–cloud environment. Extensive experiments on randomly generated DAGs with varying structural parameters (e.g., density and CCR) compare DRL-ECTS against Random, Greedy, HEFT, and CPOP baselines. Under high dependency (density = 0.9), DRL-ECTS reduces the average makespan by 17.9% compared with HEFT, and under communication-dominated settings (CCR = 1.0) it achieves a 9.66% reduction. These results demonstrate the effectiveness of learning-based server allocation for DAG scheduling in heterogeneous edge–cloud systems.