Reinforcement learning for freight booking control problems

Booking control focuses on the problem of deciding whether to accept or reject bookings to maximize revenue while considering limited capacity. For freight applications, computing the cost of fulfilling requests requires solving an operational decision-making problem which often corresponds to a mixed-integer linear program. We propose a two-phase learning-based approach that first learns to predict the objective of the operational problem, then leverages the prediction within reinforcement learning algorithms to compute the policies. The method is general and applies to different problems faced in practice. We show strong performance on two booking control problems in the literature: distributional logistics and airline cargo management.