Automated Constraint Specification for Job Scheduling by Regulating Generative Model with Domain-Specific Representation

TL;DR

This work tackles the manual bottleneck of translating heterogeneous manufacturing data into formal JSP constraints by introducing a constraint-centric architecture that regulates generative models through domain-specific representations. It articulates a three-level hierarchical DSL framework with dual operation-centric and product-flow-centric views, and couples it with an automated adaptation algorithm to tailor these DSLs to diverse production scenarios. Empirical results across complete pipelines and modular evaluations show the approach outperforms pure LLMS-based baselines in constraint abstraction, generation, and schedule grounding, while maintaining reliability and interpretability. The automated scenario adaptation further demonstrates scalability across multiple manufacturing configurations, enabling broader, more dependable deployment in smart manufacturing environments.

Abstract

Advanced Planning and Scheduling (APS) systems have become indispensable for modern manufacturing operations, enabling optimized resource allocation and production efficiency in increasingly complex and dynamic environments. While algorithms for solving abstracted scheduling problems have been extensively investigated, the critical prerequisite of specifying manufacturing requirements into formal constraints remains manual and labor-intensive. Although recent advances of generative models, particularly Large Language Models (LLMs), show promise in automating constraint specification from heterogeneous raw manufacturing data, their direct application faces challenges due to natural language ambiguity, non-deterministic outputs, and limited domain-specific knowledge. This paper presents a constraint-centric architecture that regulates LLMs to perform reliable automated constraint specification for production scheduling. The architecture defines a hierarchical structural space organized across three levels, implemented through domain-specific representation to ensure precision and reliability while maintaining flexibility. Furthermore, an automated production scenario adaptation algorithm is designed and deployed to efficiently customize the architecture for specific manufacturing configurations. Experimental results demonstrate that the proposed approach successfully balances the generative capabilities of LLMs with the reliability requirements of manufacturing systems, significantly outperforming pure LLM-based approaches in constraint specification tasks.

Figures (7)

• Figure 1: Illustration of the proposed constraint-centric architecture.(A) This panel presents a running example that illustrates the complete information flow, transitioning from two formats of procedures to a grounded production plan (depicted in the top and bottom rows). It also outlines the three modules involved: constraint abstraction, constraint generation, and schedule grounding, along with their corresponding core working mechanisms (shown in the middle row). The dsl program specified for the manufacturing scenario is highlighted as the primary driving force of the architecture. (B) This panel provides an intuitive visualization of the procedural transmission of information throughout the architecture's process. Progressing from left to right, the sequence of states includes the original procedures, fully-structured route sheet, generated constraint, jsp-solver-generated schedule, and ultimately, the production plan. The information is color-coded according to its type, such as operation name, machine, duration, and unspecified information, to enhance understanding.
• Figure 2: Illustration of the algorithms for the automated production scenario adaptation of the architecture.(A) This diagram illustrates the framework of non-parametric modeling for the automated design of semantic features within both operation-centric and product-flow-centric program view dsl. (B) This diagram depicts the framework of the EM algorithm for the automated design of syntactic features within product-flow-centric program view dsl.
• Figure 3: Results of the complete pipeline evaluation.(A) Comparison of Ours with MSL and TSL across four evaluation metrics over ten scenarios. (B) Showcases of the grounded production plans generated by Ours, MSL, and TSL, respectively.
• Figure 4: Results of the constraint abstraction evaluation.(A) Comparison of Ours-CAM with MSL-I across four evaluation metrics over ten scenarios. (B) Showcases of the fully-structured route sheets generated by Ours-CAM and MSL-I, and TSL, respectively.
• Figure 5: Results of the constraint generation evaluation.(A) Comparison of Ours-CGM with MSL-II across the evaluation metrics Constraint-Acc and Runtime-ER over ten scenarios, within the isolated version of the experiment. Results for Compiler-ER are not visualized because the error cases are minor and consistent across the baseline approaches. Consequently, we have omitted the corresponding plots for the sake of brevity. This design choice aligns with the discussion by Xiao et al. xiao2023chain. (B) Comparison of Ours-CAM-CGM with MSL-I-II and TSL-I across two evaluation metrics over ten scenarios, within the incorporated version of the experiment. (C) Gantt chart visualizations of the jsp-solver-generated schedules, derived from the specified constraints, as generated by Ours-CGM and MSL-II, respectively. The Gantt chart visualization effectively illustrates the differences in modeling fidelity of the jsp specification across three approaches. The y-axis represents the various machine types involved in production, revealing that schedules derived from the baseline approaches utilize fewer machines, primarily due to the absence of resource constraints. The x-axis depicts the total production duration, showing that the baseline approaches generate shorter time horizons, likely resulting from omissions in the specification of steps, operations, or machines---a consequence of lacking both resource and precedence constraints. The occupied area within the chart indicates the total machine occupation time. Notably, our approach's schedule contains more unoccupied areas, demonstrating that more machines must remain idle due to procedural requirements---this clearly illustrates that precedence constraints are modeled with greater precision in Ours compared to the baselines. (D) Gantt chart visualizations of the jsp-solver-generated schedules, derived from the specified constraints, as generated by Ours-CAM-CGM, MSL-I-II, and TSL-I, respectively. Applying the same criteria as (C), the jsp specification modeled by the two baseline approaches demonstrates reduced fidelity, primarily attributable to their failure to incorporate resource constraints and precedence constraints.