Optimization of return freight selection

The research “Optimization of return freight selection” is performed at HST Groep, an all-round logistics service provider located in Enschede, and is focused on the road transport department. The core problem solved in the research is “the lack of an automated and standard way to select return freight from a freight exchange”, which is transformed in the following main research question: “How can the selection of return freights from the freight exchange be automated and optimized in order to automatically find the best option available?” Analyzing the context in which the research is performed, filters or selection criteria as well as performance measures have been found. The selection criteria are used as input to the platform to reduce the possible freights to select from, while the performance measures are used to judge the suitability of a selection (where both happen manually by planners). Based on these selection criteria and performance measures, it has been concluded that the problem can be modeled as a rich variant of a well-known class of problems in literature called Vehicle Routing Problem (VRP). Literature about the VRP and its relevant extensions have been researched. Moreover, three models that closely resemble the problem within the research have been analyzed and compared, where it is concluded that none of the researched models is able to model the problem that HST faces due to its unique features. Finally, the literature study ended with a study to solving methods for these kinds of problems, where it was concluded that Adaptive Large Neighborhood Search (ALNS) is the best for the current problem. The problem, for which a model is constructed, consists of a scenario with the objective of selecting orders from a set of orders (the freight exchange) and constructing routes containing these orders for each of the vehicles in such a way that profit is maximized. For this scenario, a Mixed Integer Linear Program (MILP) has been developed. Besides the profit maximization variant, three model variants are discussed: (1) profit maximization including CO2 emissions, (2) soft time windows and (3) soft time windows including CO2 emissions. To improve the performance of the MILP, some valid inequalities have been proposed. The solution design ends with a more detailed description of the ALNS, where each element of the algorithm is defined. Experiments have been conducted to test the MILP and ALNS algorithm. In total, seven experiments are being conducted, where experiment S0 was used to determine the parameter values that are used in the MILP and ALNS. The other six experiments consist of testing the effectiveness of each model variant (experiment S1-S4), the effect of adding dynamism to the model (rolling horizon, experiment S5) and testing the ALNS against a human planner at the company (experiment S6). Each experiment is executed with several different data instances, where each data instance is a unique combination of the number of orders and the number of vehicles. Based on the experiments, it has been concluded that the MILP and ALNS are both able to make an automated selection of profitable orders from a set of orders. The ALNS seem to perform about the same as the MILP, except for the larger data instances where the ALNS is outperforming the MILP. Based on these observations, it has been concluded that the core problem has been solved. The main recommendations of the research are to implement the ALNS algorithm using a pilot study and make sure that there is acceptance by the planners.