Urban Mobility Decision Intelligence๏
An open-source benchmark platform for AI-driven urban mobility decision making โ combining spatiotemporal forecasting, multi-agent simulation, and offline reinforcement learning with reproducible evaluation.
Why this project?๏
Taxi drivers waste 30โ60% of their shift searching for passengers. In NYC alone, this means millions in lost revenue annually. This project provides a reproducible, research-grade platform for testing and comparing AI-driven repositioning strategies.
Key results at a glance๏
Strategy |
NDCG@3 |
Hit@3 |
Daily fare |
|---|---|---|---|
Hot Zone |
0.7846 |
0.5842 |
$431.21 |
Single-Step |
0.9024 |
0.8804 |
$548.77 |
Two-Step (default) |
0.9565 |
0.9714 |
$570.61 |
Two-Step vs Single-Step: +$21.84/day, paired bootstrap 95% CI [$5.00, $39.53].
Architecture๏
Pipeline: Raw TLC trips โ Data cleaning โ Demand forecasting โ Multi-agent simulator โ Policy optimization โ Benchmark evaluation
Policies: Hot Zone ยท Single-Step ยท Two-Step Horizon ยท DQN ยท Double DQN
Forecasting: LightGBM, XGBoost, GraphSAGE, GAT
Simulation: Single-driver reference rollout + finite-demand multi-agent competition
๐ Documentation
๐งญ Platform
๐ฌ Research Notes
Important boundaries๏
Simulator results are not production revenue estimates. The rollout and multi-agent simulators omit congestion, airport queues, and market feedback. See Methodology.
This is not offline RL. NYC TLC data lacks logging-policy propensities. The Q-learning extension is online Q-learning inside an estimated simulator. See DQN and Double-DQN Baselines.
Exposure risk: Two-Step strategy has 70.33% airport exposure. This saturation risk is absent from single-driver simulators.
Reproducibility๏
All results are reproducible. Reference metrics are checked into the repository. Run the full pipeline:
git clone https://github.com/caizefan34/urban-mobility-ai.git
cd urban-mobility-ai
pip install -e ".[dev,forecasting,graph,rl]"
make all
Citation๏
@software{cai2025nyc_taxi_recommendation,
author = {Zefan Cai},
title = {NYC Taxi Zone Recommendation: An Open-Source Benchmark Platform for AI-Driven Urban Mobility},
year = {2025},
url = {https://github.com/caizefan34/urban-mobility-ai}
}