Investigating Smart Traffic Signal Controllers at Signalized Crosswalks: A Reinforcement Learning Approach

At signalized crosswalks, traditional traffic control strategies such as fixed-time control and detector- actuated control are likely to be inefficient, which often adds to traffic congestion. Additionally, there exists limit amount of research investigating the "metrics" or threshold performance values of applying control strategies at crosswalks. It is likely that a "smart" control strategy can minimize delay to pedestrians and motorists at crosswalks. This paper describes a reinforcement learning approach to search for an optimal traffic signal control strategy. A value function approximation technique - deep Q-network - is adopted to facilitate the learning process. To demonstrate the effectiveness of the proposed control strategy, a case study was conducted using microsimulation. The reinforcement learning framework was compared with baseline traffic signal controllers on the basis of traffic efficiency. The results indicated that existing detector-based traffic signal controllers work well in periods with low pedestrian volume. However, after pedestrian volume reaches moderate levels, the proposed reinforcement learning framework is superior. The results also confirmed that if the appropriate sensor and control technology can lead to an optimal traffic control strategy from the perspectives of efficiency, we will have achieved a form of "smart interaction" at crosswalks, which can be a useful element of smart mobility.