MD2POS Benchmark: A Multi-Domain Scheduling Dataset for Hierarchical Reinforcement Learning

MD2POS Benchmark Dataset v1.0 This dataset supports reproducible research on the Multidimensional Dynamic Parallel Online Scheduling (MD2POS) problem in digital service supply chains. It is constructed based on the problem scenario defined in the paper titled "A Graph-Mediated Hierarchical Reinforcement Learning Approach to the MD2POS Problem in Digital Service Supply Chains" (submitted to Computers & Operations Research). The MD2POS problem extends classical parallel-machine scheduling to settings characterized by the following features: Jobs arrive dynamically over time following a Poisson process. Each job consists of a directed acyclic graph (DAG) of tasks with heterogeneous multi-dimensional resource requirements (CPU, RAM, GPU, disk). Downstream tasks are determined at runtime by the results of upstream tasks, a paradigm known as Result-Driven Dynamic (RRD) branching. This problem arises naturally in domains such as bioinformatics, video transcoding, AI fine-tuning, and financial risk computation pipelines. Dataset Content: The dataset spans four application domains: genomic sequencing, video transcoding, AI fine-tuning, and financial risk calculation. It includes five problem scales: N = 50, 100, 200, 500, and 1000 jobs. For each (domain, scale) configuration, 20 independent instances with different random seeds are provided, totaling 400 scheduling problem instances. Each instance is stored as a NumPy .npzarchive, containing the task DAG structure (edge list), software-type assignments, processing time distributions, per-task resource requirements, staggered arrival times, and RRD branching probabilities. Parameter Calibration: Simulation parameters were calibrated against real-world workflow traces (from WfCommons, MLCommons MLPerf, and the 1000 Genomes Project) and empirical execution logs collected via the Bio-Fenix monitoring system. Intended Use: The dataset is designed for evaluating algorithms—including heuristic, metaheuristic, and reinforcement learning approaches—for dynamic parallel-machine scheduling with multi-dimensional resource constraints. Usage: The dataset can be loaded directly using the accompanying MD2POSEnv Gym environment (code available at [YOUR GITHUB URL]). Each file is named in the format {domain}_{N}samples_seed{seed}.npzand can be loaded via numpy.load(file, allow_pickle=True). Field descriptions and example code are provided in the README.md file. Associated Code: DDA-HRL training framework (train.py, md2pos_env.py). License: CC BY 4.0.