Feature-based selection of carsharing relocation modes

One-way and free-floating carsharing systems must be rebalanced to achieve a high service level, and thus generate benefits for users and society. In practice, vehicles can be relocated with multiple different modes (e.g., by truck or by driving them), but a single mode is sufficient in many instances. Obviously, a single mode is preferred from a computational standpoint: The routing problems are less complex since less synchronization is necessary, and thus solve much faster. It remains an open question which features drive the decision on the best mode, and if operators can decide a priori whether hybridization of several modes is beneficial/necessary, and which modes one should hybridize among. We build a classifier based on linear regression which predicts the costs for all individual modes. The advantage of this approach is that cost estimates (i) can be used as a feature in other approaches, and (ii) allow operators to estimate the necessary budget upfront. However, cost estimates cannot be used directly to determine key drivers for modal choice. We, thus, use logistic regression and decision trees for determining the best mode. These approaches are better at determining relevant features that explain which mode is preferred in an instance. We find that the most important features to decide between modes are vehicle and truck costs per kilometer as well as their velocities, and the average number of ve-hicles that shall be relocated per day (that is, the imbalance of the system). In most instances, the decision is between driving vehicles to rebalance them and rebalance staff by biking, or loading vehicles onto a truck. Hybridization proves useful in ≈20% of all instances, and a simple rule- based classifier is able to predict correctly that hybridization is necessary in most instances