Data Driven CONWIP: Similarity Aware Admission for Job Shop Scheduling

Abstract

This work investigates a data‐driven enhancement of CONWIP (Constant Work In Process) for job shop environments characterized by high product variety and routing heterogeneity. Standard CONWIP stabilizes flow by limiting total WIP, but blind admission can induce abrupt workload swings when a freed slot is filled by a job with a radically different routing or workload. We propose and evaluate similarity‐aware admission rules (Sim-A and Sim-A‐EDD) that use real‐time job attributes) to select the candidate that best matches the profile of the job that just exited, while preserving due‐date sensitivity. Using a simulation model calibrated to OKP‐type production, we compare these rules against classical dispatchers (EDD, FIFO, Slack) across three arrival-rate scenarios. Each configuration was tested with 50 independent replications and a 95% confidence analysis after a 2000‐day warm‐up and 2000‐day measurement horizon. Results show that workload‐aware, data‐driven admission significantly reduces both the percentage of late jobs and the magnitude of lateness under moderate and heavy loads, while retaining CONWIP’s operational simplicity under light loads. The findings demonstrate a practical pathway to combine pull control with online data to deliver customization without sacrificing predictability or throughput.