Research on relieve of traffic congestion based on optimization of DQN algorithm in CamTra recognition system

This research presents a method of calculation for traffic signal plans based on the DQN (Deep Q-Leaning Network) algorithm to relieve traffic congestion in the city during rush hour. It is a very important part of our independently developed Camera Tracking System (CamTra), which has been applied in various fields such as traffic big data processing, drowning warning, factory intelligence optimization, and precise logistics simulation. With the increased number of cars, the occurrence of traffic congestion is rapidly increasing. The most important difficulty addressed in this paper is the analysis of the signal plan for tempo (speed and spacing) and green wave transmission. In addition to this, the research further optimized other accompanying factors such as action space and network structure, to achieve the best adaptability for signals. With the increase of adaptability for signal, the effectiveness of the ε-greedy strategy has also been further enhanced. The improved algorithm associates the ε-greedy strategy with the cumulative reward sequence, which realizes the adaptive control of the ε-greedy strategy and balances the relationship between exploration and performance. Finally, with the experiment of SUMO (Simulation of Urban Mobility), the performance of the enhanced DQN algorithm in single- and double-crossroads traffic signal control was verified. In comparison with the traditional static signal strategy, the enhanced DQN algorithm performed better in relieving traffic congestion.