CORE DP 2025 / 04

Transfer Reinforcement Learning for Pricing, Driver Repositioning and Customer Admission in Ride-Hailing Networks / Thomas De Munck, Jean-Sébastien Tancrez, Philippe Chevalier 

> We consider the problem of a ride-hailing platform (e.g., Uber, Lyft) that connects supply with demand over a network of locations. To this aim, the platform makes pricing, driver repositioning, and customer admission decisions. Customers are impatient and have distinct willingness to pay. Drivers can be repositioned by the platform, or can choose to relocate to other locations by themselves. We formulate this problem as a discrete-time Markov decision process and propose a transfer learning approach to find an efficient policy. Our approach first derives a rolling-horizon strategy by repeatedly solving a deterministic optimization problem. Then, two neural networks are pretrained to replicate the strategy and learn the associated value function. Finally, the policy is further improved through deep reinforcement learning (DRL). Using data from New York City, we apply our approach to networks of up to 20 locations. The results show that our approach outperforms alternative DRL algorithms and rolling-horizon strategies while reducing computation time and stabilizing learning. We also explore the interplay between pricing, driver repositioning, and customer admission, providing insights into their respective roles.