NEURAL NETWORKS AND NON-PARAMETRIC STATISTICAL MODELS: A COMPARATIVE ANALYSIS IN PAVEMENT CONDITION ASSESSMENT

Much research has concentrated on developing accurate and robust pavement condition and performance prediction models whose goals are both to assess the factors that affect pavement deterioration and to predict future pavement performance. In recent years, many authors have departed from the classical statistical approaches for model development and have worked with alternative techniques, commonly known as soft computing, that are particularly well suited for data that exhibit nonlinear properties. Based on a large European database with more than 900 test sections from 15 (European) countries, this paper complements prior research in two ways; first, it compares prediction results from three different soft computing techniques, Neural Networks, Hierarchical Tree Based Regression and Multivariate Adaptive Regression Splines, on a common database and, second, it assesses the importance of various structural, environmental and traffic characteristics on pavement condition based on these flexible computational approaches. The results show that the approaches tested provide very encouraging prediction results, especially in comparison to regression models, and that the approaches evaluate differently the factors affecting performance.