Présentation sur YouTube

Many dynamic optimization problems in urban mobility and supply chain are modeled as a sequence of independent static subproblems that must be rapidly re-optimized as new information arrives. These subproblems typically repeat over time and share many common characteristics that directly impact the efficiency of the optimization strategy. This means that subproblems solved in the past can become useful in guiding the optimization process of future subproblems. This talk presents a machine learning-based metaheuristic approach that efficiently reoptimizes autonomous ridesharing plans as new online demand arrives. The optimization approach consists of a local search-based metaheuristic that iteratively revisits previously made vehicle-trip assignments through exchanges within and between vehicles. These exchanges are performed by selecting from a pool of destroy-repair operators using a machine learning approach that is trained offline on a large dataset consisting of more than one and a half million examples of previously solved autonomous ridesharing subproblems. Computational experiments are performed on dynamic instances extracted from real ridesharing data published by Uber Technologies Inc. The results show that the proposed machine learning-based optimization approach outperforms benchmark data-driven metaheuristics by about nine percent, on average. Managerial insights highlight the correlation between the selected vehicle routing features and the performance of metaheuristics in the context of autonomous ridesharing operations.

This work is in collaboration with Dr. Mor Kaspi (Tel-Aviv University), Prof. Jean-François Cordeau (HEC Montréal), and Prof. Nikolas Geroliminis (EPFL). The full paper is accessible at https://www.sciencedirect.com/science/article/abs/pii/S1366554522002198.