Learning-based Online Optimization for Autonomous Mobility-on-Demand Fleet Control

TL;DR

The paper tackles online dispatching and rebalancing for autonomous mobility-on-demand fleets by coupling combinatorial optimization with learning in a hybrid CO-enriched ML pipeline. It casts the dispatching and rebalancing problem as a $k$-disjoint shortest paths problem on a digraph and learns arc weights via structured learning (Fenchel-Young loss) to produce anticipative online policies. Two implementations, sample-based (SB) and cell-based (CB), are developed and validated on a real-world NYC Manhattan case study, where they outperform greedy and sampling baselines, achieving up to 17.1% profit improvements and robust performance across scenarios. The framework provides scalable, decision-ready policies with polynomial-time solvability for the underlying CO problem and offers insights into policy structure, externalities, and computation times for practical deployment.

Abstract

Autonomous mobility-on-demand systems are a viable alternative to mitigate many transportation-related externalities in cities, such as rising vehicle volumes in urban areas and transportation-related pollution. However, the success of these systems heavily depends on efficient and effective fleet control strategies. In this context, we study online control algorithms for autonomous mobility-on-demand systems and develop a novel hybrid combinatorial optimization enriched machine learning pipeline which learns online dispatching and rebalancing policies from optimal full-information solutions. We test our hybrid pipeline on large-scale real-world scenarios with different vehicle fleet sizes and various request densities. We show that our approach outperforms state-of-the-art greedy, and model-predictive control approaches with respect to various KPIs, e.g., by up to 17.1% and on average by 6.3% in terms of realized profit.

Figures (20)

• Figure 1: The hybrid abk:CO-enriched abk:ML pipeline.
• Figure 2: Example of the dispatching problem modeled as a digraph (a) and a corresponding solution (b).
• Figure 3: Example of an extended digraph with predicted requests in the prediction horizon $[t+1, t^{\text{prep}})$.
• Figure 4: Extended digraph with rebalancing vertices in prediction horizon $[t+1, t^{\text{prep}})$ and capacity vertices.
• Figure 5: Example digraph (left) and the corresponding optimal full-information solution (right). The solid black paths in the optimal full-information solution represent the dispatching solution as $k$-disjoint shortest paths. The abk:SL approach learns the weights of the digraph such that the solution coincides with the solid black path of the optimal full-information solution.