كنفرانس بين المللی مهندسی صنايع و سيستم ها

صفحه اصلی / دوازدهمین كنفرانس بين المللی مهندسی صنايع و سيستم ها

Driver-Compatible Last-Mile Route Sequencing via Zone Continuity and Constraint-Guided Imitation Learning

نویسندگان :

Mojtaba Movahedinia¹ MohammadSaeed Balouchi Javaran²

1- Islamic Azad University, South Tehran Branch 2- Shahid Bahonar University of Kerman

کلمات کلیدی :

last mile delivery،route sequencing،learning to rank،imitation learning،driver aware routing،zone constrained optimization

چکیده :

Last-mile route sequencing is usually evaluated through distance or travel-time efficiency, yet historical driver routes may also reflect tacit operational preferences such as zone continuity, reduced switching, and familiar local ordering. This creates a methodological gap between spatially efficient routing and driver-compatible sequencing. Using the 2021 Amazon Last Mile Routing Research Challenge dataset, this paper examines zone continuity as an interpretable behavioral cue and studies how candidate-level imitation learning can be integrated under operational constraints. A 1,000-route validation shows that zone-constrained sequencing improves DCS from 0.288 to 0.298 (paired Δ = 0.010, 95% CI [0.004, 0.016], p < 0.001) and reduces NED from 0.942 to 0.926, while increasing spatial length as an expected efficiency-compatibility trade-off. In a learning-assisted pilot, the imitation ranker reaches 66.3% Top-1 and 91.0% Top-5 next-stop accuracy, but unconstrained learned routing remains weaker at route level. These findings support a restrained contribution: driver choices are locally learnable, but route-level compatibility is more credible when learned preferences are bounded by transparent OR constraints.