Heuristics and policies for online pickup and delivery problems

Abstract

In the last few decades, increased attention has been dedicated to a speci c subclass of Vehicle Routing Problems due to its signi cant importance in several transportation areas such as taxi companies, courier companies, transportation of people, organ transportation, etc. These problems are characterized by their dynamicity as the demands are, in general, unknown in advance and the corresponding locations are paired. This thesis addresses a version of such Dynamic Pickup and Delivery Problems, motivated by a problem arisen in an Australian courier company, which operates in Sydney, Melbourne and Brisbane, where almost every day more than a thousand transportation orders arrive and need to be accommodated. The rm has a eet of almost two hundred vehicles of various types, mostly operating within the city areas. Thus, whenever new orders arrive at the system the dispatchers face a complex decision regarding the allocation of the new customers within the distribution routes (already existing or new) taking into account a complex multi-level objective function. The thesis thus focuses on the process of learning simple dispatch heuristics, and lays the foundations of a recommendation system able to rank such heuristics. We implemented eight of these, observing di erent characteristics of the current eet and orders. It incorporates an arti cial neural network that is trained on two hundred days of past data, and is supervised by schedules produced by an oracle, Indigo, which is a system able to produce suboptimal solutions to problem instances. The system opens the possibility for many dispatch policies to be implemented that are based on this rule ranking, and helps dispatchers to manage the vehicles of the eet. It also provides results for the human resources required each single day and within the di erent periods of the day. We complement the quite promising results obtained with a discussion on future additions and improvements such as channel eet management, tra c consideration, and learning hyper-heuristics to control simple rule sequences.