Valuing Imperfect Information from Inspection and Sensing in Condition-Based Roadway Pavement Management with Partially Observable Conditions

Abstract

High serviceability of roadway pavements is crucial to well-functioning roadway networks. With time and use, the condition of these roadway elements degrades and maintenance or rehabilitation (M&R) is required to ensure high levels of serviceability. As resources are limited, prioritizing the M&R actions over time is needed. Such prioritization depends on pavement condition and each pavement segment’s contribution to the functionality of the larger roadway network. This paper investigates the potential gains from scheduling M&R actions in response to continuously updated, low-quality sensor- and intermittent high-precision inspection-based condition state information for roadway networks. The problem of determining a best M&R schedule given partially and imperfectly observed conditions and based on nonstationary stochastic condition deterioration modeling is framed as a partially observable Markov decision process, and a method based on an efficient, off-policy, actor–critic deep reinforcement learning method is proposed for its solution. This solution methodology is applied to an illustrative example network to evaluate how inspection precision and frequency influence the value of information (VoI) and whether continuously sensed data can be effective as an alternative monitoring method in the absence of inspection. The value of alternative sources of information on pavement condition state, how much to pay for it, whether it can replace inspection, and whether the efforts, training of personnel, and/or equipment needed to obtain it will pay off is investigated.