GARLIC: GPT-Augmented Reinforcement Learning with Intelligent Control for Vehicle Dispatching

TL;DR

The paper addresses the challenge of efficient, driver-behavior-aware vehicle dispatching in urban ride-hailing, where local observations poorly capture global spatiotemporal dynamics. It introduces GARLIC, a GPT-augmented MARL framework that combines hierarchical traffic state representation via multiview honeycomb graphs, a GRU-based dynamic reward capturing driving behavior, and a GPT-based policy learner with a GeoLoss to align actions with real-world geography. Key contributions include a multiview graph representation that reduces communication latency, a contrastive-learning–based dynamic reward model, and a transformer-based dispatching policy that achieves seconds-scale responses while aligning with driver incentives. Experiments on two real-world datasets demonstrate GARLIC’s ability to reduce trajectory error and improve dispatch efficiency, with notable gains on larger-scale data and robust ablations validating the value of driving-behavior integration and multiview representations.

Abstract

As urban residents demand higher travel quality, vehicle dispatch has become a critical component of online ride-hailing services. However, current vehicle dispatch systems struggle to navigate the complexities of urban traffic dynamics, including unpredictable traffic conditions, diverse driver behaviors, and fluctuating supply and demand patterns. These challenges have resulted in travel difficulties for passengers in certain areas, while many drivers in other areas are unable to secure orders, leading to a decline in the overall quality of urban transportation services. To address these issues, this paper introduces GARLIC: a framework of GPT-Augmented Reinforcement Learning with Intelligent Control for vehicle dispatching. GARLIC utilizes multiview graphs to capture hierarchical traffic states, and learns a dynamic reward function that accounts for individual driving behaviors. The framework further integrates a GPT model trained with a custom loss function to enable high-precision predictions and optimize dispatching policies in real-world scenarios. Experiments conducted on two real-world datasets demonstrate that GARLIC effectively aligns with driver behaviors while reducing the empty load rate of vehicles.

Figures (11)

• Figure 1: A vehicle dispatching scenario.
• Figure 2: The framework overview of GARLIC.
• Figure 3: The multiview graph of road networks.
• Figure 4: Different combinations of multiview graphs.
• Figure 5: Results under different settings of $\alpha$.