arXiv cs.AITuesday · May 26, 2026FREE

Low-Cost Labels, Reliable Choices: Rollout-Calibrated Hyper-Heuristics for Job Shop Scheduling

The arXiv paper "Low-Cost Labels, Reliable Choices: Rollout-Calibrated Hyper-Heuristics for Job Shop Scheduling," published on May 26, 2026, introduces a novel approach to learning-assisted hyper-heuristics for the Job Shop Scheduling Problem (JSSP). These heuristics are designed to select among various dispatching rules while maintaining the feasibility and interpretability of constructive JSSP solutions. A primary challenge in this domain is the computational expense of label generation, as each supervised label typically requires rolling out candidate rules from a partial schedule. Furthermore, ensuring the reliability of a learned selector is crucial; it should only deviate from a strong default rule when a credible gain is predicted, avoiding unnecessary switches. To address these issues, the researchers propose a selector that incorporates regret-normalized rollout labels, a contextual K-Nearest Neighbors (KNN) uncertainty estimate, and a gating mechanism. This gate activates only when the predicted improvement surpasses an uncertainty-adjusted margin, ensuring decisions are robust. The study also explored varying rollout depth and breadth to analyze the cost-quality trade-off in label generation. Evaluated on synthetic JSSP instances, the gated selector demonstrated the lowest mean Relative Percentage Deviation (RPD) among all learned selectors. It remained competitive with the best fixed dispatching rule and achieved more than an an order of magnitude reduction in Random-HH mean RPD, indicating substantial improvements in scheduling efficiency and reliability for complex job shop environments.