Reproduce “Reinforcement Learning in the Wild: Scalable RL Dispatching Algorithm Deployed in Ridehailing Marketplace” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Soheil Sadeghi Eshkevari, Xiaocheng Tang, Zhiwei Qin, Jinhan Mei, Cheng Zhang, Qianying Meng, Jia Xu
RLW is an online reinforcement-learning dispatch system for ride-hailing marketplaces that combines spatial value estimation with cancellation probabilities, smoothed rewards, standardized matching utilities, and real-time graph matching. Across offline studies and production experiments, it improved service and income metrics in several markets, including a 1.342% weighted-average income gain in A/B tests and an estimated 5.3% improvement in a composite metric after large-scale deployment.
Key insights
- Accounting for destination-dependent future driver value can outperform purely myopic matching, especially in markets with strong supply-demand imbalance and dense dispatch alternatives.
- Completion-probability modeling is central: assignments with attractive nominal rewards may be undesirable when cancellations are likely, though high-cancellation markets remain difficult for online value learning.
- Standardizing utility components reduces sensitivity to changes in fare scale and pricing policies, while the benefits of long-term optimization depend on city size, market density, and operating conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.