Adaptive Grid: Intelligence for the Roadway

Inspiration

If you’ve ever driven through Rolla, Missouri, you will have spent time sitting at a red light on Highway 63 or 10th Street while the cross-street is completely empty. In smaller municipalities like Rolla, traffic signals are typically fixed-time, operating on rigid timers based on historical studies that may be years out of date.

We wanted to bridge the gap between historical trends and real-time action. We were inspired to take the data already available, and use it to train a system that actually understands a city's unique traffic.

What it Does

Adaptive Grid is a universal optimization framework. While our case study focuses on Rolla, the engine is built to work regardless of initial location.

Universal Ingestion: The system pulls real-world road geometry for any city via OpenStreetMap. By swapping a few data files, the AI can pivot from a small town to a major metropolis.
Historical Logic: The engine ingests historical traffic reports to calculate hourly "demand multipliers." This ensures the simulation reflects actual morning rushes and late-night lulls, and is largely used for traffic flow logic.
Optimizes: Using Reinforcement Learning (RL), every intersection learns from these patterns. The signals pre-train on historical data to anticipate congestion, and then dynamically adjusts cycle lengths and green-light splits to maximize flow.

How we Built it

We maintained a modular, lightweight architecture:

The Map: We utilized OSMnx to download the road geometry and intersection data for Rolla, MO.
The Backend: A Flask-powered Python server runs the simulation logic, calculating car physics (acceleration, deceleration, and gap acceptance) 60 times per second.
The RL Engine: We implemented a "Hill-Climbing" optimization algorithm. The agent monitors the average wait time and adjusts the green-light split. The system uses a feedback loop: if the total wait time decreases after a change, the agent reinforces that behavior; if it increases, the agent adjusts in the opposite direction.
The Frontend: A reactive dashboard built with Leaflet.js visualizes traffic flow and provides a live feed of the RL agent's performance.

Challenges we Ran Into

Realistic Kinematics: Modeling human-like driving, including slowing for turns and maintaining safe following distances, was complex to code. We had to implement logic so that cars could react to what is around them.
Directional Geometry: OpenStreetMap data often lacks specific stop-sign orientations. We wrote custom logic to determine road priority based on intersection geometry.

Accomplishments that we're Proud Of

Proven Optimization: When switching from Fixed Light mode to Adaptive Learning mode, the simulation shows a measurable drop in city-wide wait times (In our Rolla demo, reducing average wait time from 13 seconds to 6 seconds). The RL actually learns the rhythm of the city.
Visual Realism: The vehicles aren't just dots; they accelerate, decelerate, and queue at lights with high fidelity, reflecting actual urban driving patterns and reacting to other cars around them.

What we Learned

We gained experience in:

Reinforcement Learning: Translating traffic flow into a reward/penalty system that an AI can optimize.
Full-Stack Integration: Connecting high-frequency Python simulations to web-based UIs via REST APIs and JSON streaming.
DevOps: Managing codebase versions with Git and self-hosting our development environment to handle the simulation's computational load.

What's Next for Adaptive Grid

Browser-Side Training: Allowing users to run the machine learning training directly in the client.
Infrastructure Testing: Allowing planners to "test" what happens if a stop sign is replaced by a traffic light, or vice versa.
Expansion: Scaling the tool to support any city across Missouri, helping planners reduce carbon emissions and get citizens home faster.