Framework for Development and Comprehensive Comparison of Empirical Pavement Performance Models

Abstract

Empirical performance-prediction models are a central part of every network-level pavement management system. In this regard, a variety of novel techniques including computational intelligence have been applied, mainly without a systematic approach to ensure compliance with principles of pavement engineering. In this study, a framework is provided for development and comprehensive comparison of alternative techniques for pavement performance modeling. As an example, several machine-learning techniques are compared in developing flexible pavement-roughness prediction models using Federal Highway Administration (FHWA’s) long-term pavement performance (LTPP) data. Three important principles of model development—maximum likelihood, consistency, and parsimony—are considered in providing a robust parameterization guideline. Variant architectures of artificial neural networks (ANN), radial basis function (RBF) networks, and support vector machines (SVM) are tested to determine the optimum parameters. Final developed models are compared through quantitative and qualitative evaluations by means of a testing database that has not been used for model development. The example comparison gives the generalized RBF network model an edge over other machine-learning techniques in predicting pavement performance. This framework can be implemented by roadway agencies to develop a robust and representative performance-prediction model for pavement management systems. Moreover, the provided framework can be used to benchmark and compare alternative modeling paradigms for specific prediction problems.