Consistent anticipatory route guidance

Abstract

Anticipatory route guidance consists of messages, based on traffic network forecasts, that assist drivers' path choice decisions. Guidance is consistent when the forecasts on which it is based are verified after drivers react to it. This thesis addresses the formulation and development of solution algorithms for the consistent anticipatory route guidance generation (RGG) problem. The thesis proposes a framework for the problem, involving a set of time-dependent variables and their relationships. Variables are network conditions, path splits and guidance messages. Relationships are the network loading map, transforming path splits into network conditions; the guidance map, transforming network conditions into guidance messages; and the routing map, transforming guidance messages into path splits. The basic relationships can be combined into three alternative composite maps that model a guidance problem. Consistent guidance corresponds to a fixed point of a composite map. With stochastic maps, RGG model outputs are stochastic process realizations. In this case, the consistency fixed point corresponds to stationarity of the RGG solution process. Numerical methods for fixed point computation were examined, focusing on approaches that are rigorous and applicable to large-scale problems. Methods included Gibbs sampling for highly stochastic maps; generalizations of functional iteration for deterministic maps; and the MSA and Polyak iterate averaging method for "noisy" (deterministic plus disturbance) maps. A guidance-oriented dynamic traffic simulator was developed to experiment with RGG solution methods. Computational tests using the simulator investigated the use of Gibbs sampling to compute general stochastic process outputs; and examined the performance of the averaging methods under different model formulations, problem settings and degrees of stochasticity. Gibbs sampling successfully generated realizations from the stationary solution process of a fully stochastic model, but entails considerable computational effort. For noisy problems, the MSA found fixed points in all cases considered. Polyak averaging converged between two and four times faster than the MSA in low or moderate stochasticity problems, and performed comparably to the MSA in other problems. Formulations involving path-level variables converged more quickly than those involving link-level variables.