An Integrated Reinforcement Learning and Centralized Programming Approach for Online Taxi Dispatching.

Area of Science:

• Operations Research
• Artificial Intelligence
• Transportation Science

Background:

• Balancing supply and demand in ride-sourcing is complex due to real-time requests and traffic.
• Existing methods struggle with the dynamic nature of large-scale congested road networks.

Purpose of the Study:

• To develop a robust and scalable approach for real-time taxi operations.
• To integrate order matching and vehicle relocation decisions for improved efficiency.

Main Methods:

• A hybrid approach combining reinforcement learning (RL) and centralized programming (CP) within a Markov decision process framework.
• RL component learns state-value functions incorporating driver experience, historical demand, and network congestion.
• CP component enables cooperative, nonmyopic decision-making for drivers under system constraints.
• Advanced temporal-difference learning with prioritized gradient descent and adaptive exploration addresses sparse rewards and sample imbalance.

Main Results:

• A simulator trained on Manhattan road networks and NYC taxi data was used for evaluation.
• The proposed integrated RL-CP approach demonstrated superior performance compared to existing dispatching algorithms.
• Significant improvements in taxi driver profits and reductions in customer waiting times were observed.

Conclusions:

• The integrated RL-CP approach offers a scalable and effective solution for real-time ride-sourcing operations.
• This method enhances operational efficiency by optimizing both order matching and vehicle relocation.
• The findings suggest a promising direction for improving urban mobility and taxi services.