AI Boosts Ride-Sharing Efficiency by 27%

Ride-sharing services like Uber and Lyft promise to cut traffic and costs, but their algorithms often make short-sighted decisions that hurt long-term performance. A new study introduces an AI method that improves ride-pooling efficiency by considering future impacts, leading to faster pickups and fewer empty miles. This advancement could make shared rides more reliable and affordable for millions of users.

The researchers developed a non-myopic reinforcement learning approach that increases service rates—the percentage of ride requests accepted—by up to 8.4% compared to traditional methods. It also reduces passenger waiting times by over 27% and in-vehicle travel times by 12.5%, while maintaining service quality. By optimizing both vehicle-rider matching and idle vehicle repositioning, the system achieves these gains without requiring more vehicles, potentially cutting fleet sizes by 25% for operators.

To create this system, the team used historical data from New York City taxi trips, simulating ride-pooling scenarios with a large fleet to ensure no requests were rejected during training. They applied an n-step temporal difference learning algorithm, a type of reinforcement learning, to predict the long-term value of assigning vehicles to specific zones and times. This involved generating simulated experiences where the AI learned from past trip patterns, updating value functions based on rewards like served rides and reduced delays. The offline phase built a lookup table of spatiotemporal values, which guided real-time decisions in online components for matching and rebalancing.

Results from testing on February 2024 NYC data show that the AI method outperforms baseline algorithms across various fleet sizes. For instance, with 1,500 vehicles, it raised the service rate to 95.5%, compared to 94.3% for myopic approaches, and cut waiting times significantly during peak hours. The system proactively repositions idle vehicles to high-demand areas, reducing rejections by up to 50% at busy times. Visualizations of learned values reveal that the AI identifies valuable zones, such as commercial districts during evening peaks, improving resource allocation.

This research matters because ride-pooling can alleviate urban congestion and lower emissions by maximizing vehicle occupancy. For everyday commuters, it means shorter waits and more reliable shared rides, while operators benefit from cost savings and higher efficiency. In cities struggling with traffic, such AI-driven systems could make public transportation alternatives more attractive, supporting sustainability goals.

However, the study notes limitations: it does not account for electric vehicle charging needs or deeper integration of matching and rebalancing, which could affect real-world deployment. Future work should explore iterative learning methods and optimize travel distances further to balance efficiency gains.