Leveraging Multi-Agent Deep Deterministic Policy Gradient (MADDPG) for Real-Time Traffic Signal Optimization
DOI:
https://doi.org/10.3126/mvicjmit.v1i2.85876Keywords:
Multi-agent reinforcement learning, MADDPG, Traffic signal optimization, adaptive controlAbstract
Kathmandu’s rapid urbanization has resulted in increasing traffic congestion attributed to 1.5 million registered vehicles in 2024, including a heterogeneous mix of motorcycles (60%), buses (15%), and non-motorized users (25%). This study explores Multi-Agent Deep Deterministic Policy Gradient (MADDPG) for decentralized adaptive traffic signal control tailored to Kathmandu’s complex traffic ecosystem. Utilizing Simulation of Urban Mobility (SUMO) simulation calibrated with local traffic volumes and You Only Look Once (YOLOv5) for vehicle detection, preliminary findings indicate MADDPG reduces vehicle delay by 30-35%, queue lengths by 25%, and emissions by 12%, while improving pedestrian safety. This ongoing work suggests MADDPG as a scalable, cost-effective solution congruent with Kathmandu’s traffic infrastructure.
