Smart Traffic Signal Control Using Reinforcement Learning and Connected Vehicle Data

Project Abstract

Smart Traffic Signal Control Using Reinforcement Learning and Connected Vehicle Data investigates a data-driven framework that redefines urban traffic management by integrating real-time vehicle data with adaptive signal control. The study addresses congestion, travel time unreliability, and emissions through a scalable system that learns optimal signal phase sequences and timing using reinforcement learning (RL) while leveraging connected vehicle (CV) data to improve situational awareness and responsiveness to dynamic traffic conditions. The abstract presents the design, methodology, experimentation, and implications of a hybrid policy that combines model-free RL with domain-specific constraints to ensure safety, fairness, and real-world deployability in heterogeneous traffic environments. The proposed framework entails a multi-agent RL architecture where traffic signal controllers act as agents negotiating green, yellow, and red phases based on local state observations (queue lengths, arrival rates, pedestrian demand) and global cues inferred from CV data such as speed, headways, and vehicle density. A novel reward shaping strategy aligns learning objectives with traffic efficiency metrics (average delay, stops, and queue spillback) and environmental considerations (emissions and energy consumption). To address the sparse reward problem and non-stationarity introduced by CV data streams, the approach integrates curriculum learning, transfer learning across intersections, and an imitation learning pre-training phase using historical signal timing data and expert policies. The system is designed to handle partial observability and communication delays by incorporating state estimation techniques and robust policy updates. We implement and evaluate the framework through a multi-scale simulation platform combining microscopic vehicle dynamics with macroscopic traffic flow models, augmented by CV data streams generated from realistic penetration rates and penetration scenarios. The study explores several RL variants, including Deep Q-Networks, Deep Deterministic Policy Gradient, and proximal policy optimization, with attention to sample efficiency and convergence stability. A comparative analysis against fixed-time, actuated, and adaptive traffic control baselines demonstrates improvements in average travel time, intersection throughput, and reliability under varying demand patterns, incident conditions, and CV penetration levels. Additionally, the research examines the marginal benefits of CV data quality, latency, and privacy-preserving aggregation schemes on control performance, offering guidelines for data fusion and real-world deployment. Security, privacy, and ethics considerations are woven into the design, ensuring that data sharing adheres to constraints and that control policies do not introduce unsafe states or inequitable service to different user groups. The results reveal that RL-driven signals, when informed by CV data, can significantly reduce congestion hotspots and smoother network-wide traffic progression, even with moderate CV penetration. The study also discusses deployment challenges, including integration with legacy infrastructure, robustness to sensor faults, and the need for continuous policy updating in evolving traffic ecosystems. The research contributes a scalable, data-informed traffic signal control paradigm that promises enhanced mobility, reduced emissions, and more resilient urban transportation networks.

Project Overview

What This Project Is About

---

A plain-language overview of using smart traffic signals that adapt in real time by learning from traffic patterns and information from connected cars. The project explores how an intelligent controller can reduce congestion, waiting times, and fuel consumption by adjusting signal timing based on evolving road conditions.

---

---

The Problem It Addresses

---

---

---

Objectives of the Project

---

1. Understand the basics of traffic signal control and why adaptability helps.
2. Learn simple machine learning ideas that can guide signal timing decisions.
3. Design a basic framework that uses real-time data to adjust signal phases.
4. Evaluate potential benefits such as reduced wait times and emissions.
5. Discuss practical deployment considerations and limitations.

---

---

What You Will Do Step by Step

---

1) Study existing signal control methods and data sources from connected vehicles. 2) Gather or simulate traffic and vehicle data. 3) Build a simple decision mechanism to change timings based on current conditions. 4) Test the approach in a small model or simulator. 5) Compare with fixed-timing signals using basic metrics. 6) Analyze results and discuss real-world limits. 7) Write up learnings and potential next steps.

---

---

Expected Outcome

A demonstrable method for adaptive signal timing that uses live vehicle data to reduce delays and improve traffic flow, along with a clear set of limitations and future improvement ideas.