Triple-BERT：我们在顺风车平台上真的需要多智能体强化学习来派单吗？

于红博

摘要: On-demand ride-sharing platforms, such as Uber and Lyft, face the complex real-time challenge of bundling and matching passengers with different origins and destinations to available vehicles, while dealing with significant system uncertainties. Due to the large number of drivers and orders, order dispatching is often tackled using Multi-Agent Reinforcement Learning (MARL). However, traditional MARL methods struggle to capture global information and lack cooperation among workers, while Centralized Training Decentralized Execution (CTDE) MARL methods suffer from dimensionality issues. To address these challenges, we propose Triple-BERT, a centralized Single Agent Reinforcement Learning method tailored for large-scale order dispatching on ride-sharing platforms. Based on a variant of TD3, our approach breaks down the joint action probability into individual driver action probabilities to handle the vast action space. To deal with the extensive observation space, we introduce a novel BERT-based network that uses parameter reuse to manage parameter growth as the number of drivers and orders increases, and an attention mechanism to capture the complex relationships among drivers and orders. Our method is validated using a real-world ride-hailing dataset from Manhattan, showing an approximately 11.95% improvement over current state-of-the-art methods, with a 4.26% increase in served orders and a 22.25% reduction in pickup times. Our code, trained model parameters, and processed data are publicly available at the repository https://github.com/RS2002/Triple-BERT.

更新时间: 2025-12-31 05:05:23

领域: cs.LG,cs.AI,cs.MA

下载: http://arxiv.org/abs/2510.03257v2