CART Algorithm: A Data-Driven Approach to Automate Maintenance Selection in Pavement Management Systems

Abstract

One of the essential components for developing a plan to manage pavements is the maintenance selection step, which is done using decision trees (DTs). In most cases, DTs are developed based on experts’ judgment. These trees are subjective and result in inconsistent decisions over time. A more objective approach is to use the data-driven trees which was addressed in multiple research studies. However, the resulting trees were limited in terms of the prediction accuracy levels achieved, the types of pavements covered, and the predicted maintenance action types. The goal of this paper is to use pavement management system (PMS) data to improve the consistency of the decision-making process in PMS to make more accountable decisions. To achieve this goal, this paper’s main objectives are to utilize the classification and regression trees (CART) algorithm to create a data-driven maintenance selection DT model and to develop a data-driven model impact evaluation approach to assess the benefit and cost of using the data-driven DT. An extensive dataset covering the three pavement types in the state of Iowa is used to achieve this objective. Also, a possible approach to evaluate the impact of using data-driven DTs to select maintenance actions instead of the subjective DTs is investigated. With 10 internal nodes and 52.5% overall prediction accuracy of five maintenance action types, the final DT model developed in this paper showed an improvement in the prediction accuracy and model complexity compared with the trees presented in the literature. This model is also expected to save the agency an average of $3.9 million compared with using the PMS output.