Hybrid seminar at GERAD or Zoom.
A coffee break will be served before. Please register here to attend the event.

This activity will take place as part of the thematic semester on sustainable mobility.
This activity is followed by “Last-Mile Logistics via Robot-Assisted EV Routing and Customer-Centric Pickup: A Two-Stage Matheuristic Framework”.

In ridesharing platforms, system efficiency relies heavily on the assumption that drivers will follow optimal or predicted paths. However, drivers exhibit heterogeneous and context-dependent preferences—often deviating from navigation suggestions due to familiarity, safety concerns, or personal habit. Standard assignment algorithms that rely on "point estimates" of driver behavior fail to capture this variance, leading to inaccurate arrival time predictions and poor on-time performance. This presentation introduces a contextual inverse optimization approach to improve rider-driver assignment by learning the full distribution of driver routing preferences. Rather than assuming a single "most likely" route, our pipeline learns a model mapping contextual features to a rich class of preference distributions using score-function gradient estimation. We then solve a risk-averse assignment problem that accounts for variations in driver preferences. In a synthetic ridesharing setting, this approach predicts driver choices with 1.7–11\(\times\) lower error than leading baselines. Crucially, it reduces "post-decision surprise"—the discrepancy between planned and actual service times—by up to 24\(\times\), demonstrating that accounting for driver autonomy is key to delivering a consistent user experience.