TokenJSP: Graph Representation Learning Datasets for Job Shop Scheduling

TokenJSP: Tokenization Approaches for Attention-Based Reinforcement Learning for Solving Job Shop Scheduling Problems [Link] The dataset contains graph-based state representations of randomly generated Job Shop Scheduling Problem (JSP) instances. It was created for experiments on graph representation learning, graph autoencoders (GAE/VGAE), graph neural networks, and reinforcement learning-based scheduling. For each JSP instance, multiple intermediate scheduling states are collected by performing random rollouts in a disjunctive graph JSP environment. The environment uses left-shift scheduling whenever possible, normalized observation spaces, matrix-based graph observations, and task-level action selection. Each collected state contains: - A: adjacency matrix of the current disjunctive graph state with shape (T, T), where T = number of jobs × number of machines. - X: node feature matrix with shape (T, F), containing the scheduling status of each operation and additional critical-lower-bound-based graph features. - M: machine correspondence matrix with shape (T, number of machines), representing the machine assignment of each operation. The node features include a binary scheduled-operation indicator and CLB-based structural features computed from the current graph state and base operation features. The base environment features consist of one-hot machine assignments and normalized processing-time information. The datasets are stored as compressed NumPy archives (.npz) and contain the following arrays: - A: stacked adjacency matrices, shape (N, T, T) - X: stacked node feature matrices, shape (N, T, F) - M: stacked machine correspondence matrices, shape (N, T, number_of_machines) - n_jobs: number of jobs - n_machines: number of machines Here, N denotes the number of collected graph states, not only the number of original JSP instances. Since several states are collected from each random rollout, each generated JSP instance contributes multiple training samples. The datasets were generated for different JSP sizes, including 6x6, 10×10, 15×15, and 20×20 problem instances. They were used to train and evaluate graph autoencoder models that learn latent operation-level graph representations for downstream reinforcement learning experiments. The dataset is intended for research on: - graph representation learning for scheduling, - graph autoencoders and variational graph autoencoders, - tokenization methods for attention-based reinforcement learning, - neural combinatorial optimization, - job shop scheduling problems.