Balancing supply and demand for ride-hailing: A preallocation hierarchical reinforcement learning approach

Abstract

Ride-hailing platforms have revolutionized the travel experience for passengers. However, a fundamental problem in these platforms is the imbalance between supply and demand, especially in hot and cold regions. Most existing studies on fleet management to address this issue are based on combinatorial optimization and reinforcement learning, which focus on capturing the spatial-temporal relationship between current supply and demand while ignoring potential demand. In this paper, we propose a novel approach to ride-hailing fleet management based on preallocation hierarchical reinforcement learning (PHR), which can integrate traffic demand prediction and vehicle relocation. PHR decomposes the ride-hailing fleet management problem into two sub-problems, namely demand prediction and vehicle relocation. And then, we develop a multi-view spatial-temporal convolution module for potential demand prediction and a hyper-parameter self-attention preallocation module for vehicle relocation. Substantial experiments based on real data from multiple cities show that PHR provides superior performance in terms of platform revenue and order response rate in fleet management tasks.