Reinforcement Learning Traffic Optimization for Smart Cities

Urban traffic congestion remains a critical challenge to the development of sustainable, ef-ficient, and responsive smart cities. Conventional traffic signal systems, such as fixed-time and actuated methods, often struggle to cope with real-world traffic changes. Fixed-time signals follow predefined schedules, while actuated signals use sensors to detect cars and adjust the lights accordingly. However, both approaches can be slow to react to sudden traffic build-ups or unpredictable conditions, leading to longer delays and congestion. We have investigated the use of reinforcement learning (RL) for adaptive traffic signal control, with a specific focus on the influence of reward function design on system performance. Three reward strategies have been developed and evaluated using the Simulation of Ur-ban Mobility (SUMO) platform, namely (1) minimizing vehicle waiting time, (2) reducing queue length, and (3) a weighted-sum approach of both. Simulation scenarios include isolated intersections and connect nodes in urban network models to mimic real-world situations. Simulation results demonstrate that the designs of reward functions have ob-servable impacts on performance in handling fluctuations in traffic flow. The combined reward did not simply balance one goal against another. Instead, it helped the agent re-duce vehicle delays in real time while also keeping queues from becoming too long, espe-cially between the two intersections. This meant the controller could respond to changing traffic without letting the queue stretch past the space available between signals, which would block the upstream intersection and cause further delays. System adjusts to traffic as it happens, rather than making a trade-off between reducing delay and clearing con-gestion. These findings highlight the importance of incorporating physical constraints in-to the design of reward functions in RL-based traffic signal controllers for intelligent transport systems in urban environments.