Preference learning for efficient bundle selection in horizontal transport collaborations

To improve routing efficiency, transport service providers can enter a horizontal transport collaboration that uses a combinatorial auction to reallocate delivery orders. To find the optimal allocation, the carriers have to report bids for all possible combinations of available delivery orders. As this number grows exponentially with the number of orders to be reallocated, they are faced with an enormous computational challenge. To lift this burden, the auctioneer may offer only a limited set of order combinations. However, selecting this limited set is itself a stochastic combinatorial optimization problem known as the Bundle Selection Problem. In contrast to previous one-shot approaches to solve this problem, in this paper, a partial preference learning scheme is applied that iteratively queries carriers’ valuations, uses their responses to train preference models and then uses these fitted models to estimate valuations for new combinations of orders. This work investigates different ways to realize such a concept and analyzes their respective improvement in collaboration gains. The results indicate that the suggested algorithm can yield travel time savings of up to 20% higher than those achieved by a random benchmark and up to 10% higher than those of a literature benchmark if at least 40 query-response pairs are considered. • Auction mechanisms of horizontal carrier collaborations are information-inefficient. • We propose a preference learning approach to capitalize on available information. • Iterative query-response process elicits training data to fit regression models. • Case study with a 3-carrier collaboration network in the city of Vienna, Austria. • Linear regression models perform best, beating a random and a literature benchmark.