v3.0.0 · Research Platform · MIT License

Urban Mobility Decision Intelligence

An open-source research platform for AI-driven fleet repositioning. Demand forecasting, multi-agent simulation, offline reinforcement learning, and reproducible policy evaluation — all in one system.

Launch Demo Star on GitHub Documentation Paper Draft

0.9565

NDCG@3 (Two-Step)

3,360 queries

+$139.40

Daily Revenue Lift

vs Hot Zone baseline

+$53.74

DQN Advantage

per driver / day

402

Tests Passing

CI verified

263

NYC Taxi Zones

full coverage

Platform Capabilities

What This Platform Does

A complete decision intelligence pipeline — from raw trip data to deployable policies with rigorous evaluation.

NYC TLC Data → Demand Forecasting → Supply-Demand Model → Multi-Agent Sim → Policy Optimization → Evaluation (OPE) → API / Dashboard

Demand Forecasting

GraphSAGE + temporal attention predicts per-zone demand at 15-min granularity. Two-step retrieval achieves 0.9565 NDCG@3 across 3,360 queries.

GraphSAGE NDCG@3

Multi-Agent Simulation

263 NYC taxi zones modeled as a supply-demand network. Drivers reposition based on learned policies; revenue and idle time tracked per episode.

263 Zones Episode-based

Offline Reinforcement Learning

IQL and DQN trained on historical trajectories. Doubly-robust OPE validates policies without online deployment risk.

IQL DQN OPE

Reproducible Evaluation

402 passing tests, deterministic seeds, and published benchmarks. Every result is reproducible from raw TLC data to final metrics.

402 Tests CI/CD

Benchmark Results

Policy Performance Comparison

DQN achieves the highest daily revenue lift per driver while maintaining competitive service rate.

Policy	Daily Revenue/Driver	Service Rate	Idle Time	Advantage
Hot Zone (baseline)	$247.30	68.2%	41m	—
Random Reposition	$198.40	55.1%	58m	-$48.90
IQL	$281.50	74.6%	32m	+$34.20
DQN BEST	$301.04	76.8%	28m	+$53.74

Results averaged over 30-day simulation on NYC TLC data (Jan 2024). Revenue in USD per driver per day.

Quick Start

Get Running in 60 Seconds

Install, configure, and launch your first simulation.

install.sh

# Install the platform
pip install urban-mobility-ai

config.py

from mobility_ai import Config

# Point to your local NYC TLC trip data
config = Config(
    data_path="data/nyc_tlc/",
    api_key="YOUR_API_KEY",
    zones=263,
    horizon="15min",
)

config.save("config.yaml")

run.sh

# Launch a 30-day DQN simulation
python -m mobility_ai simulate \
  --policy dqn \
  --days 30

Start Building Smarter Mobility

Clone the repository, run the demo, and join the research community.

Launch Demo Read the Docs