Intelligent Traffic Signal Management using DRL for a Real-time Road Network in ITS

The acceleration of urbanisation and the development of the pace of industrialisation help to grow the population of metropolitan areas, thus increasing the density of traffic flow. The sole way of managing traffic congestion is to mitigate it through optimising traffic signals at the intersections of a vast road network. The synchronization amongst the traffic signals at intersections is strongly needed in order to alleviate congestion and to allow vehicles to travel smoothly along intersections. Reinforcement Learning (RL) techniques in Intelligent transportation system (ITS) are not feasible for the management of traffic signals of large road networks due to enormous information of the state-action pairs. To overcome this problem, the emerging technology of Deep Learning allows RL to form Deep Reinforcement Learning (DRL) to measure up previously unwavering decision-making issues, for handling high-dimensional states and action spaces. DRL agents perform tasks through perception, monitoring the environment through action and learning as well as analysing the results of actions. In the present work, a single DRL agent is trained using the Policy Gradient algorithm in four different categories of Deep Neural Networks (DNN) to control the traffic signals dynamically. In case of a static road network, the functional implementation and efficacy of the Policy Gradient algorithm cannot be analysed accurately due to the less intricate details of static network. Hence, two different dynamic real time road networks have been considered here. Moreover, the real-time spatio-temporal information congregated from the dynamic real time map is provided as an input, so that the traffic signal duration can be adjusted adaptively in order to manage the traffic flow appropriately. The overall success of the agent’s efficiency in different DNN models is compared here using simulation experiment. The viability of the simulation experiment is investigated using three separate simulation metrics against the baseline, which is fixed signal duration frameworks and indeed the suggested method outperforms the baseline. Moreover, The GRU model excels all other models in both the dynamic networks.