anticipatory freight scheduling in synchromodal transport
ArturoEduardo Perez Rivera is a PhD student in the research group Industrial Engineering and Business Information Systems (IEBIS). His supervisor is prof.dr. J. van Hillegersberg from the faculty of Behavioural, Management and Social sciences (BMS).
Over the last decade, Logistic Service Providers (LSPs) have explored new approaches to control their operations with the objective of improving their network-wise efficiency and reducing their environmental impact. One of such approaches is synchromodal transport. In synchromodal transport, the choice of mode, the choice of transport path, and the timing of operations is not fixed up front, but decided upon various moments using the latest information about the status of the multi-modal network and about freight demand. This increased flexibility provides LSPs with more opportunities for consolidation and options for efficient transport, throughout the network and throughout time. However, to achieve such gains, transport decisions must consider their consequences in the entire network and anticipate on their effect in future decisions and performance. In this dissertation, we study how to make such decisions in synchromodal transport and what improvement opportunities can be expected.
Our work deals with anticipatory scheduling decisions for the transport of freight in a synchromodal network. These decisions are dynamic, i.e., change when new information is revealed, and consider current and future freight demand and performance. We use four different perspectives for decisions on a multi-modal network, based on traditional scheduling methods considered in literature. For each perspective, we develop mathematical models and heuristic algorithms that support anticipatory scheduling decisions. Furthermore, for each perspective, we evaluate the output of our models and algorithms using simulation-based experiments, and provide insights into their efficiency gains, over traditional scheduling methods, using different network characteristics. These perspectives are as follows.
First perspective: We study the scheduling of long-haul transport in a network without transfers (e.g., train in a corridor, barge in a deep-sea port). We model the problem as a Markov Decision Process (MDP) and design an Approximate Dynamic Programming (ADP) algorithm to solve the model heuristically. We design different ADP architectures and provide insights into the design and evaluation challenges of using ADP in our problem. We demonstrate the effect of various network characteristics on the cost savings of using our approach instead of single-period optimization. The savings range between 6% and 9%, on average in 7 out of the 8 instance categories tested, and up to 26% in some cases. Our approach performed best with unbalanced demand and a majority of pre-announced orders.
Second perspective: We study the scheduling of long-haul transport in a network with intermediate terminals and transfers. We extend the MDP model of the first perspective and use mathematical programming techniques to capture the space-time evolution of freights in the network. We extend the ADP algorithm with additional components from Bayesian exploration, specifically the Value of Perfect Information (VPI), to handle the complex relation between performance over time and the space-time evolution of freights in the network. We describe how the one-step look-ahead perspective of traditional ADP can make the algorithm flounder and end in a local-optimum, and how through the quantification of a value of exploration in VPI, the ADP algorithm can escape this local-optimum and improve the solution. We show how our proposed ADP-VPI combination, with a heavy restriction on the decision space, achieves more than 20% gains over the profit achieved by common practice unrestricted heuristics. When considering the decision space restrictions as part of the problem, the gains are larger than 50% in 6 out of 9 instance categories tested. The distribution of time-windows was the key characteristic influencing the gains of our approach in our experiments.
Third perspective: We study the scheduling of drayage operations considering initial terminal flexibility and a categorization of jobs. Drayage operations, also known as pre- and end-haulage, are a special case of the vehicle routing problem where timing, routing, and long-haul terminal assignment decisions are integrated and simultaneously considered. We develop a Mixed-Integer Linear Programming (MILP) model of the problem and design a series of valid inequalities and pre-processing mechanisms to reduce its complexity. We design two matheuristics to solve the MILP model in a static and dynamic way. These matheuristics iteratively confine the solution space of the MILP using several adaptations, and based on the incumbent solutions, guide the subsequent iterations and solutions. We benchmark our approach against a heuristic from literature and observe cost savings between 3% and 4% in 3 out of the 4 instance categories tested. Our approach performed best with clustered locations and short time-windows.
Fourth perspective: We study the integrated scheduling of drayage and long-haul transport. We combine the models and heuristics from the second and third perspective, and design an iterative simulation approach that outputs a unified decision policy (i.e., function for anticipatory scheduling that considers the current status of the network and future performance) for scheduling drayage operations and long-haul transport. The approach from the second perspective yields estimates of future long-haul costs, which are used to assign terminals to incoming pre-haulage freights in the approach of the third perspective. We show that our approach results in cost savings between 4% and 38% in all instance categories tested. Our methods achieved the largest savings with a balanced destination distribution.
As an addition to the scheduling methods, we study how to increase the awareness about the trade-offs considered by our methods and how to facilitate the adoption of the algorithms from the first perspective through a serious game. We develop a web-based game that educates on the benefits of anticipatory decisions and demonstrates how our algorithms can support an LSP planner. We verify and validate our game design in gaming sessions with students of Industrial Engineering and Management at the University of Twente.
Finally, we present a closing reflection about our anticipatory scheduling methods for freights in synchromodal transport, and provide an outlook for further research with respect to extensions of the models, improvements of the heuristics, and implementation aspects.