Design and evaluation of real-time adaptive traffic signal control algorithms

Abstract

This dissertation investigates methods of real-time adaptive traffic signal control in the context of single isolated intersection and coordinated urban network applications. A primary goal in this dissertation is to identify and address scenarios where real-time optimized controllers do not maintain competitive performance with off-line calibrated, vehicle-actuated control techniques. An extensive literature review is supplemented by subsequent simulation experiments. Several strategies were implemented and evaluated, including OPAC, PRODYN, COP, ALLONS-D, Webster's optimized fixed-time control, and vehicle-actuated control. In particular, evaluation is based on simulation of a single, isolated intersection, where all algorithms are required to adopt the exact, deterministic traffic model used by the simulation. This approach eliminates confounding factors in comparison of algorithms, such as detector placement and disparate traffic models, focusing evaluation on the efficiency of the algorithms and their ultimate performance in terms of vehicle delay. A new algorithm is developed, employing neuro-dynamic programming techniques, also known as reinforcement learning techniques. Several very effective pruning strategies are also constructed. The final product is a very efficient algorithm capable of solving problems up to 2000 times faster than the most efficient previously published algorithm tested, with an 8% decrease in delay. This algorithm is then extended to a generalized, multi-ring control formulation. Simulation results with a standard dual-ring, eight-phase controller demonstrate that efficient, real-time solutions are achieved with a corresponding 12--22% reduction in delay relative to dual-ring, vehicle-actuated control. The real-time optimized, multi-ring controller is finally extended for urban network applications, expanding the objective function to consider downstream performance measures, and adopt standard, vehicle-actuated type coordination constraints. Control on an 8-intersection arterial is evaluated using a CORSIM simulation over a range of traffic conditions. Results are compared with TRANSYT optimized fixed-time control, coordinated vehicle-actuated control, and RHODES. Two regimes of control are revealed, where cyclic coordination constraints provide a significant benefit, and where they prevent more effective control. An adaptive coordination layer is prescribed as a unifying architecture with the potential of obtaining effective control under both regimes. The adaptive control layer specification is explicitly distinguished from existing algorithms, such as SCOOT, SCATS, and VFC-OPAC.