AI-Powered Traffic Orchestration: Reinforcement Learning at Traffic Intersections

A smart traffic light can respond to real-time traffic conditions. Systems grounded in reinforcement learning (RL) are poised to enhance signal timings with the goal of mitigating congestion and enhancing the overall transportation experience. Unlike earlier research, our analysis not only consider vehicle's speed but also take other parameter like velocity of vehicle and distance of vehicle from intersection. Further these parameters are considered for states of reinforcement learning algorithm. Actions are decided on the base of states and taken according to ACT variable which instruct further for phase change. A new reward function is designed with linear combination of waiting time and queue length. DQN algorithm is implemented, in order to evaluate two parameter namely reward and waiting time for each episode which signifies how well RL agent perform. Traffic light based on reinforcement learning are broadly investigated in the literature, are premeditated on various parameter like objective, input parameter, output metrics, methodology, and experiment work. The training period of the model was completed with simulator SUMO. Results signifies that after certain episodes reward function performs better. Proposed approach is compared with fixed time methodology and the performance comparison records a significant improvement in terms of average waiting time.