A Comparative Study on a Dynamic Pickup and Delivery Problem : Improving routing and order assignment in same-day courier operations

Sammanfattning: Pickup and Delivery Problems (PDPs) constitute a class of Vehicle Routing Problems (VRPs) consisting of finding the optimal routes for a fleet of vehicles to deliver requests from a set of origin locations to a corresponding set of destinations. PDPs are NP-hard and have a wide variety of variants and potential constraints. This thesis evaluates methods for solving a dynamic single- vehicle PDP restricted by multiple time-related constraints. The problem is dynamic in the sense that new requests arrive as time is simulated and inserted into the vehicle’s pickup and delivery plan as it is being executed. The time- related constraints include limited time windows during which the requests may be picked up or delivered, as well as maximum ride times that items may spend in the vehicle before being delivered. To solve the problem, we adapt insertion heuristics based on Large Neighborhood Search (LNS) and Heuristic Destroy and Repair (HDR) to the problem and evaluate them in a comparative study. Solution methods for the PDP are also applied on the problem of dynamically assigning incoming orders to vehicles in a delivery service with a short planning horizon. A PDP-based order assignment strategy is compared with assignment strategies based on proximity and workload. Due to the short planning horizon of the target application, the study is focused on finding well-performing methods for quickly solving small PDPs containing 10-15 requests. Our results indicate that LNS outperforms HDR for small problem instances. However, the quick convergence of HDR allows it to outperform LNS for larger problem instances. We also show that applying a PDP- based assignment strategy in the order assignment problem allows the service to accommodate more requests than the alternative assignment strategies while simultaneously providing a significant reduction in operational costs. Future work may improve the order assignment strategy by incorporating more anticipatory functionality and streamlining the PDP methods with more efficient tests for the feasibility of solutions.