Kenneth D. Kuhn

Network-Level Infrastructure Management Using Approximate Dynamic Programming

This research introduces the use of approximate dynamic programming to overcome a variety of limitations of distinct infrastructure management problem formulations. The form, as well as the parameters, of a model specifying the long-term costs associated with alternate infrastructure maintenance policies are learned via simulation. The introduced methodology makes it possible to manage large heterogeneous networks of facilities related by budgetary restrictions and resource constraints as well as by dependencies in maintenance costs or deterioration. In addition, the methodology is particularly well suited to consideration of multiple types of infrastructure condition data at the same time, including continuous-valued data and relevant historical data. Introduced techniques will prove valuable when high-quality deterioration and cost estimation models are available but are ill suited for use in a Markov decision problem framework. Computational studies show that the introduced approach is able to find an optimal solution to a relatively simple infrastructure management problem, and is able to find increasingly good solutions to a more complex problem.

Pavement Network Maintenance Optimization Considering Multidimensional Condition Data

Pavement management systems inventory historical and current conditions of roadway networks, predict the future conditions of such networks, and suggest schedules for maintenance, repair, and rehabilitation activities. Such systems typically rely on a composite condition index, a one-dimensional and often discrete measure of the overall structural health and/or serviceability of pavement. The index is used during deterioration modeling, user and agency cost estimation, and selection and scheduling of maintenance activities. Pavement can suffer from a large number of related but distinct distresses. Difficulties associated with unobserved heterogeneity have hampered efforts to accurately model deterioration through composite condition indexes. At the same time, optimization techniques used to generate recommended maintenance plans have been shown both to be sensitive to deterioration model specification and to become computationally intractable as condition data increase. This research describes how a large network of related sections of pavement, each one of which may be plagued by a number of different distresses, can be managed without reducing condition data to a composite index. The use of approximate dynamic programming mitigates the curse of dimensionality that has haunted distinct Markov decision problem formulations of the maintenance optimization problem and limited their complexity. A computational study illustrates how the proposed approach leads to more sophisticated maintenance decision rules, which can be used to ensure the suggestions of pavement management systems more closely match engineering best practices. The use of multidimensional condition data can also yield more accurate deterioration models and cost estimates. The techniques introduced in this paper in the context of pavement management could easily be applied within any infrastructure management system.

Traffic Flow Forecasting and Spatial Data Aggregation

Research investigated ways to forecast traffic flow over an area covering multiple links of a roadway network. The question arose of whether to construct one large-scale model or to combine results from separate smaller-scale models. Intuition favors link-specific analyses, but econometric results show how using aggregate models can increase accuracy. It is shown how theory holds that correlations between traffic conditions on different roadway links and errors in link-specific observations of traffic conditions increase the relative accuracy of larger-scale models, whereas site-specific sensitivities suggest smaller-scale data. Empirical evidence is presented regarding the relative accuracy of various models of traffic flow rate. One form of a seasonal autoregressive integrated moving average time series model for general flow forecasting is chosen, along with methodologies for automatically selecting models based on input data. Studies based on data from the Tokyo Metropolitan Expressway show that increasing data aggregation consistently increases model accuracy. Sums of forecasts of link flows yield predictions of areawide flow roughly as accurate as predictions based on univariate analysis of areawide data. Areawide flows are easier to predict accurately than link flows. Modeling large-area traffic flow with a univariate model is considerably simpler but produces less subsequently useful results than use of separate, link-specific models.