Urban Ridesharing: Improving Tradeoffs Between Geographic Fairness And Efficiency
Date:
Abstract: State-of-the-art order dispatching algorithms for ridesharing use deep reinforcement learning techniques along with ILP-based optimization to achieve high efficiency in terms of service rates (proportion of passenger requests accepted). However, in pursuit of efficiency, such approaches might lead to disparity in service rates based on geographic locations, leading to some passenger groups being under-served. In this talk, I present some methods that allow us to improve group-level geographic fairness in such ridesharing systems. We propose an online method that leverages the ILP-based structure of the problem to incorporate fairness into existing matching algorithms without any retraining. Interestingly, we find that it is possible to significantly improve geographic fairness with minimal loss of overall service rate when using our methods.