Book Description
Stochastic traffic and mobility simulation models are popular tools for modeling urban transportation networks. However, their use for optimizing urban transportation networks can be challenging due to their computationally intensive nature. This thesis focuses on high-dimensional simulation-based (SO) optimization problems. To find solutions with good performance efficiently, we need to balance exploration and exploitation. We propose techniques for achieving a better balance between exploration and exploitation when tackling high-dimensional SO problems in urban transportation. The first part of the thesis considers a general-purpose exploration mechanism and introduces exploitation components to it. We propose an inverse cumulative distribution function (cdf) sampling mechanism that makes use of problem-specific prior information in the form of an analytical model to efficiently sample for points with good performance. The inverse cdf sampling mechanism can be used in conjunction with any optimization algorithm. We study whether problem-specific prior information should be used in the exploration (i.e., sampling) mechanism and/or the exploitation (i.e., optimization) algorithm when tackling a high-dimensional traffic signal control problem in Midtown Manhattan. The results show that the use of inverse cdf sampling mechanism as part of an optimization framework can help to quickly and efficiently identify solutions with good performance. The second and third parts of the thesis focus on developing a framework to enable high-dimensional Bayesian optimization (BO) for stationary and dynamic transportation SO problems respectively. BO naturally combines exploration and exploitation. In the second part, we consider stationary problems and propose approaches to incorporate problem-specific prior information in the BO prior functions such as to jointly enhance both exploration and exploitation. This is done through the use of a stationary analytical surrogate traffic model. In the third part, we extend the BO framework to tackle dynamic problems by formulating and embedding a computation ally efficient dynamic analytical surrogate traffic model. For both parts, we evaluate their performance with a traffic signal control problems for a congested Midtown Manhattan (New York City) network. The proposed methods enhance the ability of BO to tackle high-dimensional urban transportation SO problems.
