Application Of Probabilistic Graph Neural Network In Traffic Signal Control

In recent years,China’s urbanization process is accelerating,and more and more cities are facing serious traffic congestion.With the progress of artificial intelligence technology,researchers hope to solve the current problem of urban traffic congestion and build an efficient urban traffic system by studying urban intelligent traffic signal control.Recent studies have used reinforcement learning to solve the cooperative control problem of traffic signals at multiple intersections.However,the reinforcement learning modeling of existing research methods not only does not consider the uncertainty and timing characteristics of multi intersection complex traffic conditions,but also these reinforcement learning methods lack theoretical support.In addition,the reinforcement learning model applied to traffic signal control has some disadvantages,such as high training cost,slow model convergence,and so on.To overcome the limitations of the above traffic signal cooperative control methods,the thesis proposes a new model: traffic signal control based on probabilistic graph neural network（TSC-GNN）.Based on uncertainty modeling,the model innovatively combines variational inference technology with graph attention networks.By describing the posterior probability distribution of intersection characteristic hidden variables,the expectation and variance are calculated to represent the uncertainty of traffic conditions.In addition,TSC-GNN applies the attention mechanism to the feature extraction structure of the recurrent neural network（RNN）,so that the traffic signal control model can dynamically learn the timing features of multiple adjacent intersections.Finally,aiming at the problem that the current traffic signal modeling based on reinforcement learning lacks a theoretical basis,the thesis design the reinforcement learning state of the TSC-GNN model by combining traffic engineering theory and queuing theory,and explains the rationality of state design in reinforcement learning modeling,to avoid the heuristic modeling of reinforcement learning.Aiming at the problems of long training cycle of deep reinforcement learning,overfitting of environment and poor generalization ability of the model,the thesis innovatively apply meta-learning to traffic signal control at multiple intersections,introduces the idea of meta-learning based on TSC-GNN model,and proposes Meta-DQN model（meta deep learning Q network model）.Through three parts: basic meta-learning,generalized meta-learning,and prior knowledge transfer,the algorithm realizes the prior knowledge of initialization parameters based on a set of tasks,and finally improves the generalization ability and model effect of TSC-GNN model,and speeds up the convergence speed of the model.Consequently,to prove the availability of the model method proposed in the thesis,the thesis chooses Cityflow traffic signal simulation platform and secondary develops the core class.For the TSC-GNN model,the thesis fully evaluated the model on real traffic data sets and synthetic traffic data sets and explained the model through ablation experiments.For the Meta-DQN model,the thesis appraises the model according to its convergence rate and model training time on three real traffic data sets and three synthetic traffic data sets.The final experimental results show that the performance of the proposed model is improved by about 8% compared with the existing baseline model.