Optimization-based Iterative Learning for Anticipatory Traffic Signal Control

In urban traffic networks, delays at signal-controlled interactions are a main component of the travel time experienced by road users. Effective optimization of traffic signal control offers high opportunities to reduce delays and overall to improve traffic operations. A strategic network traffic signal control policy, also called anticipatory network traffic control, determines signal timings to optimize network-wide objectives, e.g. minimize network travel times. While optimizing control variables, anticipatory traffic control explicitly takes into account road users’ route choice responses and the resulting network flow patterns. A traffic assignment model is usually used for approximating this route choice response. The basic problem in implementing anticipatory network control in a real-life system is that: due to the inherent model-reality mismatch of this approximated response model, the true resulting traffic pattern will in general not be optimal; signal control using incorrect anticipation might even trigger undesired traffic conditions, for instance unexpected delays, traffic congestion and spillback. Hence, since a model will unavoidably represent a coarse approximation of reality, there is a need for introducing corrective measures to deal with the model-reality mismatch and in turn achieve an effective control. The general aim of this PhD research is to develop control methods that elevate the traffic system to its best achievable performance, by explicitly addressing the model-reality mismatch. We take advantage of the fact that routine traffic patterns repeat from day to day, allowing one to learn from observing the true traffic patterns, resulting from a response to control, to compensate for errors in the response model. The solution method is initially inspired from industrial applications in robot manipulator, which have a similar objective, i.e improving performance of unknown / uncertain systems that operate repeatedly. A rule-based Iterative Learning Control (ILC) is applied to the traffic signal setting. The operational signal timings are iteratively corrected to achieve the predefined desired traffic state. The basic idea behind the iterative learning approach is to learn from the errors between observations and the desired reference. Following this basic idea of learning from observed errors, an optimization-based iterative learning approach is elaborated in the anticipatory control context. By making use of the traffic assignment model, the desired state is no longer predefined, but rather endogenously optimized. A method of Iterative Optimizing Control with Model Bias Correction (IOCMBC) is then proposed. IOCMBC generates a sequence of control settings based on a first-order model bias correction. Control performance is improved by learning from the model bias between real measurements and model predictions. It is proven that IOCMBC has two desirable properties: convergence in real traffic operations, and consistency between the convergent point and the real optimum. An integrated framework is then presented to jointly perform IOCMBC and model parameter estimation. The objective is to complete the iterative optimizing control scheme in which the underlying network flow modeling is imperfectly calibrated. Two objectives are fulfilled by this integrated framework, i.e. better control for the real system, and better model calibration. This integrated approach is advantageous to applications in which the underlying model is not only used for designing optimal control, but also for supporting other traffic management measures, for instance traffic information or route guidance. In order to link the methodology to general network applications, the key algorithmic implementation issue regarding the calculation of flow sensitivity is later addressed. We present a method to estimate the derivative of real flows with respect to signal control variables; this derivative information is particularly important in choosing directions to improve control optimization objectives. The derivative estimation method obviates the need to explicitly explore sensitivity by adding perturbations on each individual control variable, and hence allows for multiple control variables. By an evaluation on the derivative estimation, two types of estimation errors are revealed, i.e. truncation error and noise error. The presence of measurement noise is particularly detrimental to the estimation, which ultimately affects the quality of control solutions. A reliable dual method, i.e. dual IOCMBC is then proposed, improving both optimization objective and derivative estimation during the control process. Dual IOCMBC is able to generate a control sequence that converges to the real optimal point, and meanwhile forms a basis for estimation of the real flow derivatives. Numerical experiments on a general midsize network present an important intermediate step towards field implementations.status: publishe