Adaptive optimization models for infrastructure management.

摘要

We study the problem of developing maintenance and repair policies for civil infrastructure networks under model uncertainty. Model uncertainty refers to uncertainty in the choice or estimation of models to represent deterioration. We present adaptive optimization formulations for both facility and network level problems. The formulations require the choice and estimation of a set of deterioration models that can be combined to obtain representations of facility deterioration. Model uncertainty is captured by including probability mass functions over the set of models in the state-space of the problem. The probability mass functions represent beliefs about which combination of models provides an adequate representation of deterioration, i.e., which model governs the process. The probability mass functions are updated based on periodic observations of condition. This results in a representation of deterioration that changes dynamically.; We present closed-loop and open-loop-optimal-feedback control formulations for the facility-level problem, and compare them to the existing approach to maintenance and repair decision-making based on a single model, static representation of deterioration. We show that the benefits of adaptive control policies increase as model uncertainty increases or as the initial error in estimation of deterioration increases. A comparison of the two adaptive control policies reveals situations where there is additional value in applying maintenance and repair actions that produce information that reduces model uncertainty. This illustrates the probing-optimizing dichotomy in infrastructure management and validates the choice of methodology. The open-loop-optimal feedback control formulation is then extended to the network level problem. The formulation can account for interactions of the facilities that comprise the network. The formulation provides a basis to generate condition-model dependent policies that account for heterogeneities in the network. This is relevant in infrastructure management because heterogeneities are often produced by unobservable factors. We show that the benefits of adaptive control policies increase as the heterogeneity increases.; In the second part of the dissertation, we present Temporal-Difference learning methods for maintenance and repair decision-making without a deterioration model. This can correspond to a case of extreme model uncertainty where data to choose and estimate deterioration models are not available. Temporal-Difference learning constitutes an approach to maintenance and repair decision making that is radically different than the existing approach. We conduct a simulation study that shows that the methods are promising as an alternative to the existing approach, and can therefore be used to assess the costs and benefits associated with generating data to model deterioration.