Factorized Scheduling Principle: Learning Interpretable and Transferable Policies via Structured Additive Functions

Scheduling problems arise from repeatedly selecting one item from a set of candidates based on their states. These problems often reduce to assigning priority scores and choosing the highest-ranked item. In this work, we propose a factorized scheduling principle (FSP) framework to learn interpretable and transferable scheduling rules. The FSP framework represents system states as condition distributions and decomposes a global scheduling principle into additive univariate and pairwise components with identifiability constraints. The scheduling principle enables the framework to maintain a simple priority-based structure during deployment. This principle is learned by using a policy-based objective combined with a temporal-difference signal defined on the condition distribution. Experiments on synthetic and realistic scheduling tasks demonstrate the FSP framework's strong performance, interpretability, and zero-shot generalization across different system scales.

Scheduling systems are essential in many real-world settings, where decisions often involve repeatedly choosing one item from many candidates. We started this research because many learning-based scheduling methods can perform well in the exact setting where they were trained, but they are often difficult to understand and may need to be retrained when the number of candidate items changes. To address this problem, we introduce the Factorized Scheduling Principle, a method that learns a reusable priority rule from experience. Instead of learning a black-box policy tied to a fixed list of candidates, our method learns how each item’s features affect its priority. The learned rule is decomposed into feature effects and feature interactions, so users can inspect which factors make an item receive higher or lower priority. Our results show that such a learned priority rule can remain effective when applied to systems with different numbers of candidate items, without retraining. This matters because practical scheduling systems often change over time, and rebuilding a new policy for every system configuration can be costly or impractical. Our work suggests a practical step toward learning explicit priority rules from experience while retaining the transparency and reusability associated with priority-based scheduling.