Secure Operation Research On Complex Traffic Networks Based On Several Machine Learning Algorithms

Traffic networks are a type of complex networks that,together with commercial networks,social networks,power networks etc.,sustain an important part of the activities of our technological era.Secure operation of traffic networks is crucial to guarantee a good service level.This thesis focuses on two important aspects of secure operation based on several machine learning algorithms:accurate traffic light management to avoid traffic congestion and cope with different levels of traffic flow,and reliable realtime traffic predictions to provide useful information to car users and traffic managers.Unfortunately,in the current state of the art,secure operation of these two aspects is far from reality.For example,traffic light management is performed inside traffic centers under the supervision of traffic engineers.Unfortunately,no matter how experienced the personnel in the traffic center could be,it is clear that they cannot operate the traffic light infrastructure securely in an optimal way under any circumstances.For this reason,traffic light management often relies on simple rule-based strategies that do not reach secure operation.For traffic prediction,the state of the art is as follows:standard methods for supervised learning(such as deep neural networks used for traffic prediction in large traffic networks)require a long training time in order to provide reliable prediction.As a result,in case the traffic prediction needs to be reconfigured to cope with a new traffic situation(e.g.road works,maintenance etc.),it is impossible to provide in real-time a new reliable prediction.This thesis proposes new methodologies and algorithmic tools for secure operation of traffic networks:we proposed new adaptive dynamic programming methods for traffic light operation,and new supervised learning tools for traffic prediction.More specifically,the main reserach contents of this thesis can be summarized as follows:First,this work concerns a switching-based control formulation for multi-intersection and multi-phase traffic light systems.A macroscopic traffic flow modelling approach is first presented,which is instrumental to the development of a model-based and switchingbased optimization method for traffic signal operation,in the framework of ADP.The main advantage of the switching-based formulation is its capability to determine both’when’ to switch and ’which’ mode to switch on without the need to use the cyclebased average flow approximation typical of state-of-the-art formulations.In addition,the framework can handle different cycle times across intersections without the need for synchronization constraints and,moreover,minimum dwell-time constraints can be directly enforced to comply with minimum green/red times in each phase.Simulation experiments on a multi-intersection and multi-phase traffic light systems are presented to show the effectiveness of the method.Second,Switch-based ADP is an optimal control problem in which a cost must be minimized by switching among a family of dynamical modes.When the system dimension increases,the solution to switch-based ADP is made prohibitive by the exponentially increasing structure of the value function approximator and by the exponentially increasing modes.This work proposes a distributed computational method for solving switch-based ADP.The method relies on partitioning the system into agents,each one dealing with a lower-dimensional state and a few local modes.Each agent aims to minimize a local version of the global cost,while avoiding that its local switching strategy has conflicts with the switching strategies of the neighboring agents.A heuristic algorithm based on consensus dynamics and Nash equilibrium is proposed to avoid such conflicts.The effectiveness of the proposed method is verified via traffic test cases.Third,a fast architecture for real-time(i.e.minute-based)training of a traffic predictor is studied,based on the so-called Broad Learning System(BLS)paradigm.The study uses various traffic datasets by the California Department of Transportation,and employs a variety of standard algorithms(LASSO regression,shallow and deep neural networks,stacked autoencoders,convolutional and recurrent neural networks)for comparison purposes:all algorithms are implemented in Matlab on the same computing platform.The study demonstrates a BLS training process two-three orders of magnitude faster(tens of seconds against tens-hundreds of thousands of seconds),allowing unprecedented real-time capabilities.Additional comparisons with the Extreme Learning Machine architecture,a learning algorithm sharing some features with BLS,confirm the fast training of least-square training as compared to gradient training.Fouth,we propose the first hybrid recursive implementation of BLS(which we call HR-BLS).This new implementation,while being equivalent to the standard BLS in terms of trained network weights,it can smoothly train much larger networks than the standard BLS(the standard BLS can result in "out-of-memory" failures,whereas the proposed one can train larger and large networks).This means that BLS is projected toward the big-data frontier.Finally,this paper discusses the relationship between the proposed research and complex networks security from the perspectives of complex networks,cybernetics and control.