查看文章

By recent advances in technology and Artificial Intelligence (AI), traffic signal control systems 12 are preferably designed to have intelligence rather than rule-based structure. Deep 13 Reinforcement Learning (RL) as a solution for sequential decision-making problems has been 14 extensively used for adaptive traffic signal control (ATSC) systems. The deep RL-based ATSC 15 systems have shown promising results versus current actuated (rule-based) ATSC systems. The 16 conducted studies have employed different data resolutions either collected in vehicular 17 network (eg, location and speed of individual vehicles) or from camera devices (eg, queue 18 length, density, average speed) for proposed models. However, the impact of different data 19 resolutions on deep RL-based ATSC systems training performance has not been studied yet. 20 In this study, we compare the three different data resolutions in terms of computation time, 21 training stability and results for variety of performance measurements. The Double Deep Q-22 Network (DDQN) algorithm is utilized as our intelligent agent. To test and evaluate the 23 different data resolutions, a real isolated intersection is modelled in a simulation environment 24 with real traffic volume demand. The experimental results have shown that vehicular network 25 high resolution data can only contribute to a slight improvement versus camera data in terms 26 of reduction in travel time, queue length etc. at the expense of more computation time in 27 training models. Also, the camera data is more accessible compared to vehicular network data 28 which needs sensors on plenty of vehicles in network …