Research On Urban Traffic Light Control Method Based On Deep Reinforcement Learning

Since the 21 st century,the global population has been increasing.The increase of urban population has also seriously affected the development of urban transportation.The development and realization of smart transportation has become an urgent problem for all countries.With the development of the Internet and technology,new scientific and technological fields such as big data,cloud computing,deep reinforcement learning,and artificial intelligence have become the hotspots and difficulties of research.The research of intelligent transportation has also moved toward the adaptive development stage based on new technology.The realization of more efficient urban traffic light control algorithm has become an important and research value.This paper optimizes the existing deficiencies of urban traffic light control algorithms.The main contents are as follows:(1)Firstly,how to model the urban traffic light control is described in detail,and then the improvement of the storage of the Q table state space explosion in the single intersection traffic light control algorithm and the impact of the historical strategy on future learning are improved: The near-end strategy framework considers the impact of the execution strategy of the single-action historical action phase on the current execution phase,and optimizes the current intersection environment learning rate and deep learning sampling by the ratio of the current strategy to the historical strategy,in order to prevent the ratio from appearing in the super-domain.The problem is solved by using a confidence interval method.The effectiveness of the algorithm was verified by experiments in mild traffic flow and heavy traffic flow environment.(2)At present,although there are modeling schemes involving multi-junction cooperation in the study of urban traffic light control algorithms,it does not reflect how the specific cooperation between neighbors is,This paper uses the Distributed deep Q network to realize urban multi-road cooperative control modeling.The main consideration is the historical state of the intersection itself,the historical phase action,the traffic state of the previous moment of the one-hop neighbor intersection of the intersection,the influence of the phase action on the current intersection,and the historical state of the intersection itself-the action and the neighbor of the hop.The state-action sequentially calculates the phase action to be performed at the current moment of the current intersection through the MLP neural network,the LSTM neural network,and the MLP neural network.The effectiveness of the algorithm was verified by experiments in mild traffic flow and heavy traffic flow environment.(3)Urban traffic light control system belongs to distributed system.At present,there are few algorithms for distributed multi-junction multi-strategy in urban traffic light control algorithm,It is impossible to efficiently solve the problem of priority driving of vehicles of related types in cities,This paper uses distributed W-Learning to realize distributed multi-junction multi-strategy coordination of traffic lights.Control algorithm.The Q value execution corresponding to the maximum W value is selected mainly by calculating the Q value(execution phase)and the W value(importance weight)and the C value(coordination coefficient)of the intersection local strategy and the remote strategy,wherein the collaboration is based on the principle of cooperation diagram Parallel exchange of status,action,reward,W,etc.between the intersections,reducing system learning time and understanding the traffic conditions at neighboring intersections.The effectiveness of the algorithm was verified by experiments in mild traffic flow and heavy traffic flow environment.