Research On Intelligent Traffic Signal Control Under Road Network Model

With the successive increase of vehicle ownership in modern urban areas,traffic congestion has gradually become one of the most pernicious and persistent problems affecting modern urban development.Waste of resources,pollution emission and time cost caused by traffic congestion not only lead to huge economic loss,but also seriously affect people’s quality of life,which has gradually become a long-term social problem that is difficult to be cured.Traffic management is gradually becoming a worldwide problem and challenge.In recent years,the development of deep reinforcement learning has proposed new possibilities to study such problems: by reasonably setting the reward function,the intelligent body interacts with the environment and learns the traffic signal control strategy,so as to intelligently control the traffic signal to achieve the goal of reducing the average travel time and ultimately alleviating traffic congestion.However,most of the existing studies have not fully considered the complexity and variability of realistic urban road networks,and there are still many problems that need to be solved.Based on the research at the frontier of traffic signal control,this thesis will propose several optimization measures for the deep reinforcement learning model in the road network road environment,verify the feasibility and superiority of the model in terms of cooperative control through simulation experiments,and explore the selection of the optimal parameter set for the reinforcement learning model in the road network environment through extensive experiments.The main research work of this theis is as follows:(1)This thesis proposes a deep reinforcement learning traffic signal control model based on road network pressure,which uses a normalized pressure value optimized for the road network model to define a deep reinforcement learning reward function,and verify through simulation experiments that the indicator can effectively characterize the intersection congestion in the road network environment.(2)This thesis proposes a traffic signal control model for road networks based on Dueling Double Deep Q Network,using three improvement measures to optimize the traditional Deep Q Network model,based on the improvement ideas of Dueling Deep Q Network and Double Deep Q Network to alleviate the Q value overestimation problem.A prioritized experience replay strategy is also introduced to help the model train and converge faster.The optimization measures is verified through simulations.(3)The thesis proposes a Multi-agent cooperative traffic signal control algorithm based on graph attention network for a realistic road network environment where the influence between neighboring intersections is constantly and dynamically changing with the road conditions.This thesis uses graph attention network with multi-head attention mechanism to aggregate feature information of intersections and their neighbors to realize the cooperative interaction of each intersection.It is experimentally verified that the attention score can effectively represent the influence of each neighborhood intersection on the central intersection.(4)Based on the above multiple optimization measures,the intelligent traffic signal control fusion model is proposed.This thesis builds a simulation experiment system based on fusion model.The superiority of the fusion algorithm under the road network model is verified through comparative experiments.Moreover,the selection of the state space and reward function of the reinforcement learning model in the road network environment is explored in depth,and the optimal parameter combination definition is obtained through a large number of comparative experiments.