Technical University of Crete Adaptive Fine - Tuning for Large - Scale Nonlinear

Despite the continuous advances in the fields of control and computing, the design and deployment of an efficient Large-scale Nonlinear Traffic Control System (LNTCS) remains a significant objective. This is mainly due to the complexity and the strong nonlinearities involved in the modeling of traffic flow processes. Practical control design approaches are often based on simplified models about the system dynamics, leading to LNTCS with suboptimal performance, as the use of more complex models of effective LNTCS is virtually unavoidable in most complex control system applications. The ultimate performance of a designed or operational LNTCS (e.g. urban signal control, or ramp metering) depends on two main factors: (a) the exogenous influences, e.g. demand, weather conditions, incidents, and (b) the values of some design parameters included within the LNTCS. When a new control algorithm is implemented there is a period of, sometimes tedious, fine-tuning activity that is needed in order to elevate the control algorithm to its best achievable performance. Fine-tuning concerns the selection of appropriate (or even optimal) values for a number of design parameters included in the control strategy. Moreover, the continuous mediumand long-term variations of the traffic system dynamics call for a frequent or even continuous maintenance of LNTCSs. When an operational but “aged” control algorithm needs to be updated the same fine-tuning procedure has to take place, which – if done properly – is extremely costly. Typically, this fine-tuning procedure is conducted manually, via trial-and-error, relying on expertise and human judgment and without the use of a systematic approach. Currently, a considerable amount of human effort and time is spent for initialization or calibration of operational LNTCSs, which does not always lead to a desirable outcome. In many cases, the result is that system maintenance is neglected and the system performance deteriorates year after year. This thesis introduces and analyzes a new learning/adaptive algorithm that enables automatic fine-tuning of LNTCS, so as to reach the maximum performance that is achievable with the utilized control strategy. The proposed Adaptive Fine Tuning (AFT) algorithm is aiming at replacing the conventional manual optimization practise with a fully automated online procedure. The thesis provides a detailed analysis of the algorithm as well as a stepby-step application description. Finally, application results of the algorithm to real-time fine-tuning problems of general LNTCS are presented. The efficiency and online feasibility of AFT algorithm is investigated through extensive simulation experiments for two LNTCS. The first test case is a large-scale ramp metering control problem. A multivariable ramp metering regulator is applied to the stretch of the