Research On Adaptive Control Algorithm Of Isolated Intersection Based On Vehicle Arrival Characteristics

With the explosive growth of social progress,economic development,and urban travel demand,people have put forward more stringent requirements for the experience and comfort of traffic travel,while the urban infrastructure construction is far from meeting the growth of travel demand,and the urban congestion problem is becoming more and more normalized.In the face of such problems,many scholars have researched traffic signal optimization control to solve the problem of urban traffic congestion.The traditional signal control mode has limitations from the current practical application.In the face of multi-mode complex traffic flow,the performance is not ideal,and the solution can not be given in real-time and long term.With the development of intelligent technology,the optimal control of traffic signals is gradually developing in the direction of ’ model-free and self-learning ’.Researchers have studied the optimal control method of traffic signals in combination with reinforcement learning and other technologies and achieved good results,which provides the possibility for achieving more efficient traffic signal adaptive control schemes.This paper starts with analyzing the vehicle arrival characteristics at the signal-controlled intersection and explores the traffic state by constructing a traffic state evaluation model.At the same time,the adaptive control model is established by using reinforcement learning theory,and the adaptive control method of intersection traffic signals is studied.The specific work is as follows.1.The identification index of vehicle arrival characteristics is established.Based on the three traditional traffic evaluation indexes of occupancy,speed,and flow,this paper introduces the entropy weight method and mathematical set pair method to construct the two indexes of the Smooth Index and Level Index proposed in this paper.This paper explores the sensitivity of the lifting index to the change in vehicle arrival characteristics and the feasibility of the feature description from the aspects of determining the weight coefficient and the time interval of data acquisition.The feasibility of the two indexes to describe the arrival characteristics of vehicles is verified by an example.2.An intersection adaptive control model based on Q learning is constructed.Under the relevant constraints,the discrete state space and the reward function more sensitive to the traffic control effect are established by combining the characteristics of the value learning algorithm model.The feasibility and applicability of the level index as an element of constructing state space are verified by experiments.3.A deep reinforcement learning algorithm model based on an unblocked index is constructed.The DDPG algorithm model is built based on the time distribution characteristics of traffic flow arrival characteristics and the characteristics of the strategy learning algorithm model.In the VISSIM simulation road network model,the feasibility of the control model is verified by the actual road environment and data.