Optimizing traffic flow with Q-learning and genetic algorithm for congestion control

Traffic congestion in urban areas presents significant challenges, adversely affecting economic productivity, public health, and overall quality of life. Efficient coordination of traffic signals emerges as a crucial strategy to mitigate these impacts. This paper introduces an innovative approach to traffic management by leveraging Q-learning and Genetic Algorithms (GAs) to optimize traffic light schedules, aiming to reduce vehicle waiting times at intersections. The approach proposed in this study is implemented in a sophisticated simulation environment, facilitated by the python-traffic simulator platform, leveraging real-time data. Uniquely, in this paper, Q-Learning implementation incorporates a novel yet redundant random shuffling of action values in the value determination process, which differs from standard Q-learning approaches. Through a comparative analysis, we evaluated the performance of these advanced methodologies against the default traffic light control behavior. The proposed algorithm demonstrated a substantial improvement, reducing average vehicle waiting time. The research thoroughly assesses the performance of simulation outcomes under various scenarios, examining episodes in batches of 20, 50 and 100. The method exhibits notable improvements over traditional traffic control algorithms. It reduces the average wait time by approximately 12.54% compared to the default fixed cycle method. Also showcases a significant reduction in the average wait time by approximately 10.39% with the second method (longest queue first). In comparison to the third method (search algorithm) the proposed method demonstrates an appreciable decrease in the average wait time by approximately 6.09%. These findings underscore the potential of applying machine learning and evolutionary computation techniques to enhance traffic flow efficiency, suggesting a scalable solution for urban traffic management challenges.