Vehicle routing with time windows and driver learning

This research investigates the problem of route construction for local deliveries and pickups of packages with time windows, and accounting for driver learning. An objective of this research is to develop realistic yet efficient optimization models for strategic route planning and daily vehicle dispatching with stochastic customer locations and demand, with time window constraints due to service priorities. Meanwhile, the dispatching plan maintains driver familiarity with their service territories. There are two objectives in route construction. The first objective is to maximize the driver familiarity. Through repetition of visiting service areas, drivers become familiar with their service territory and their performance improves. The more frequent drivers visit their territories, the more efficient the driver dispatching plan becomes. The second objective is to construct vehicle dispatching plans that are sufficiently flexible to serve stochastic customer locations and demand.; Therefore, to optimize the tradeoff between these two objectives, the research proposes a two-stage vehicle routing model: Strategic Core Area Design (SCAD) and Operational Cell Routing. In the SCAD model, we propose the concepts of "cell", "core area" and "flex-zone" to cope with large-scale vehicle dispatching problems and the benefit of driver familiarity. Two SCAD methods are proposed as alternatives. The first method is the single core area for drivers to serve their assigned cores all day and the second allows drivers to serve the delivery cores and pickup cores separately. The results of core areas from the SCAD stage then are used in a daily operational routing model to construct daily routes.; 452 instances were tested and provide results that have shown that both core area methods reduce the total duration up to 5% and the driver utilization by about 5%. In the R type problem, the SCAD1 outperforms SCAD2 when the learning limit is less then 70% and SCAD2 outperforms SCAD1 when the learning limit is greater than 70%. The separate core method provides less number of drivers, total duration, and waiting time than the single core method in the RC type problem, which is a mix of business customers and regular customers. The RC type problem represents a more realistic problem that consists of business customers, which are clustered over the area, and regular customers, which are uniformly distributed over the area. Our strategic solutions lie within 6% above a lower bound and daily routing builds promising routes. Relative to the daily route construction plan, our method significantly reduces the computational time, especially with large problem sizes and also maintains the consistency of the dispatching route.