Business Analytics (BA) Seminar: Jan Fabian Ehmke, University of Vienna

Speaker: Jan Fabian Ehmke, University of Vienna (https://busan.univie.ac.at/members/univ-prof-dr-jan-fabian-ehmke/)

Title: Preference learning for efficient bundle selection in horizontal transport collaborations

Abstract: 

To reduce routing costs and emissions in the home delivery sector, transport service providers (carriers) can collaborate by reallocating their delivery orders among them, such that all collaborators gain some profit compared to the setting in which they operate in isolation. Due to a lack of mutual trust, carriers prefer to keep the level of information sharing to a minimum. Combinatorial auctions have proven to strike a balance between high collaboration gains and limited information sharing as they account for the substitution and complementary interactions between the delivery orders that shall be reallocated. At the heart of such combinatorial auctions is the Winner Determination Problem (WDP). This optimization problem computes the order-to-carrier assignment with the lowest overall routing costs. Thus, all carriers must report for each possible combination of orders the marginal costs of including them into their existing routes. However, as computing this marginal cost entails solving the Vehicle Routing Problem with Time Windows (VRPTW), solving the WDP in its entirety is too demanding and time consuming for practical applications. A remedy to this situation is presented in the Bundle Selection Problem (BuSP), which limits the number of bundles that carriers must bid on with as little decrease in collaboration gain as possible. In this paper, we generalize existing approaches from the transportation literature and apply methods from the field of preference learning that are new for solving the BuSP in horizontal carrier collaboration. In particular, we consolidate the steps of bundle selection and bidding into a unified iterative process to increase the use of available information for decision-making. To that end, we demonstrate (1) the necessity of taking the WDP constraints into account already in the bundle selection stage and (2) that fitting machine learning models to predict carrier bids can aid the auctioneer in selecting better bundles. Our results indicate that this preference learning scheme can outperform the conventional “one-shot” approach to bundle selection.

Host: Sanne Wøhlk

Organisers: Surabhi Verma and Hartanto Wong