Adaptive Signal Controller based on Convolution Neural Network Agent for Heterogeneous Traffic Environment: A Case Study in Shah Alam, Malaysia

Abstract

This paper presents an adaptive traffic signal system based on the convolution neural network (CNN) technique. This work differs from others by evaluating the adaptive controller in a simulated stochastic traffic environment and liberating the signal operations to an acyclic assignment. In addition, a practical communication protocol is integrated to capture state conditions. The agent optimises signal operation using a management policy for waiting time gaps between competing travel directions. The testing models included morning and evening peak hours, and the performance of the CNN system is compared against a fixed time controller representing the status quo of the studied isolated junction. The findings indicated that the CNN control significantly improved the travel time between 25% and 35% compared to the existing control method. The work also includes sensitivity analysis to determine the impact of the time plan and saturation flow level on the performance of the proposed control system.