📞 +91-7667918914 | ✉️ iarjset@gmail.com
International Advanced Research Journal in Science, Engineering and Technology
International Advanced Research Journal in Science, Engineering and Technology A Monthly Peer-Reviewed Multidisciplinary Journal
ISSN Online 2393-8021ISSN Print 2394-1588Since 2014
IARJSET aligns to the suggestive parameters by the latest University Grants Commission (UGC) for peer-reviewed journals, committed to promoting research excellence, ethical publishing practices, and a global scholarly impact.
← Back to VOLUME 13, ISSUE 11, NOVEMBER 2026

Coordinated Multi-Agent Reinforcement Learning for Urban Traffic Simulation

Mirabai Girase, Manoj V. Nikum, D. M. Patel

👁 3 views📥 1 download
Share: 𝕏 f in ✈ ✉
Abstract: Traffic congestion at signalized urban intersections causes life-threatening transit delays for Emergency Vehicles (EVs) such as ambulances and fire appliances. Traditional emergency signal preemption systems depend heavily on isolated optical, acoustic, or induction loop sensors. When an approaching emergency vehicle is detected within a local vicinity, these systems forcibly end active conflicting phases to secure an immediate green light. However, such isolated and greedy preemption disregards corridor-level progression and downstream queue states, generating catastrophic secondary queue spillbacks and prolonged recovery delays on intersecting commuter avenues. In this paper, we introduce Emergency Corridor Intelligence (ECI), a coordinated Multi-Agent Reinforcement Learning (MARL) framework built on Centralized Training with Decentralized Execution (CTDE) and parameter-shared Deep Q-Networks (MADQN). Each intersection agent observes local phase durations, approach queue distributions, and emergency vehicle telemetry, enhanced by spatial state representations from neighboring junctions along the corridor.
The framework evaluates the proposed MARL controller against both conventional Fixed-Time cyclic control and Actuated Rule-Based Greedy Preemption across Low (300 veh/hr), Medium (700 veh/hr), and High (1200 veh/hr) demand regimes in Eclipse SUMO microscopic simulations under strictly matched random seeds (S = 42). Experimental results show that while greedy preemption cuts Emergency Vehicle Travel Time (EVTT) by 35.0%–36.8% at the cost of massive peak queues (up to 39 vehicles on cross-streets), the proposed MARL controller successfully resolves the trade-off curve—attaining EVTT reductions of up to 25.7% while dramatically lowering average commuter delay by 63.7% to 80.9% and reducing peak network queue lengths by more than 50%. The system includes an internal Safety Transition Arbiter that enforces mandatory 3-second yellow change intervals and minimum green times, ensuring compliance with legal requirements, no collisions, and no conflicting green signals. This project provides a reproducible, academically validated,and fully operational end-to-end platform that includes real-time TraCI telemetry streaming, interactive dashboard visualization, and automated PDF performance certification.

Keywords: 1. Multi-Agent Reinforcement Learning (MARL): Parameter-shared Deep Q-Networks that enable decentralized traffic control across multiple intersections.

How to Cite:

[1] Mirabai Girase, Manoj V. Nikum, D. M. Patel, “Coordinated Multi-Agent Reinforcement Learning for Urban Traffic Simulation,” International Advanced Research Journal in Science, Engineering and Technology (IARJSET)

Creative Commons License This work is licensed under a Creative Commons Attribution 4.0 International License.