||Title of Dissertation: Intelligent Classification, Simulation and Control of Traffic Flow Fengxiang Qiao, Doctoral of Philosophy, 2000 Dissertation directed by: Asso. Prof. H. Yang Department of Civil Engineering The Hong Kong University of Science and Technology Most of physical traffic models, which can explain many of the traffic phenomena and are widely used in the analysis, management, control and guidance of traffic flow, are normally in well-posted mathematical forms. However due to the complexity of traffic properties, physical models in some cases can not really match the depicted actual phenomena even though certain correcting factors are introduced. Hence it is hoped to develop intelligent models that can auto-adapt, or self-calibrate from the input-output data in the applications. Although some pioneer works on these intelligent data dependant models have been reported recently, the researches on these kinds of models are far fewer than required. This dissertation aims to apply a series of intelligent and robust approaches to some of the problems in the classification, simulation and control of traffic flow. At first, the efficient, yet practical, method for the classification of traffic flow states on highways was developed. The method sampled actual flow data for each possible case, recognized their different characteristics, and then sorted them into various clusters using neural network pattern recognition techniques. A small-scale test with actual data was conducted, and the method was found to be potentially applicable in practice. Then, a simulation approach was proposed to analyze the occurrence of traffic conflicts at unsignalized intersections. The simulation model can provide some useful statistics on traffic situations at unsignalized intersections, which are also useful for safety assessment and examination of the necessity for traffic control devices such as road signs or traffic signal controls. After that, a neural network based system identification (NNSI) approach was used to establish an auto-adaptive model for simulating and forecasting the dispersion of traffic flow on road segments. The structure and linking weights of the neural network based model could be on-line calibrated, and thus it could be applicable for operations under varying traffic environments. Simulation and field validation results provide strong evidence of the good performance for the proposed neural network based system identification approach. Next, the fuzzy logic based delay and performance index estimation systems and the corresponding intersection timing techniques were proposed, together with simulation study and field studies. The fuzzy logic based delay estimation approach, which can combine the complex technical and non-technical factors, can be adaptively suitable to the changing environment and easily implemented. The fuzzy performance index system can give a general index for the optimization of a signal timing plan at an intersection. The rule base of the fuzzy performance index system came from the psychology survey at the intersection concerned. This approach can be used not only in isolated intersection timing but also in arterial signal timing, network signal timing, traffic guidance systems etc. Wide applications can be expected in the field of traffic control engineering. Finally, the robust control theory was introduced into the traditional freeway ramp metering system. Using the data dependent system approach, the freeway plant can be identified the on-line system operation with no predefined knowledge on the flow-density relationship, while the sub-optimal robust controller is constructed to reduce some kinds of uncertainties possibly existing in the control system. The simulated results show that the controlled freeway density can be maintained around the desired target within a small error band converging quickly. Widely promising applications can be found in freeway management systems.