Bayesian Approach to Updating Markov-Based Models for Predicting Pavement Performance

Abstract

The Markov decision process is one of the most common probabilistic prediction models used in infrastructure management. When existing data are insufficient, expert knowledge is commonly used to derive a Markovian transition probability matrix. Eventually, every pavement management system will progress to a level at which inspection measurements from the network will be organized into a database to be used for performance prediction. The best way to use this body of data to improve the initially developed transition probability matrix is to combine prior expert knowledge with new observations. This paper proposes a method for periodically updating Markovian transition probabilities as new inspection data become available. Bayesian inference is used to accommodate uncertainty in the expert-derived initial probabilities and measurement errors from inspection of the network. A data set of asphalt concrete pavement observations from the Minnesota Department of Transportation test facility is used to illustrate the proposed method.