Prediction of Highway Bridge Performance by Artificial Neural Networks and Genetic Algorithms

Bridge management systems (BMS) comprise various techniques need to help make decisions on the type of works that need to be performed to maintain the serviceability of a bridge and to extend its useful life. These decisions rely on current and future bridge conditions therefore it is essential for a BMS to accurately predict the future bridge performance, or in other words to assess the extent of bridge deterioration. Numerous deterioration models are reported in the literature. Most of these methods were developed using probabilistic approaches ranging from Markovian methods to regression techniques with various levels of detail. While offering mostly marginal improvements, such methods increase the complexity of the procedures and level of expertise needed. Besides, high reliance of these methods on historical data, which are likely to contain missing information, reduces the chances for a reliable model. The ability of learning in Artificial Intelligence (AI) methods provides promising results in modeling and forecasting even in the existence of non-linear complex relationships. Furthermore, easier use of AI tools provided by today?s software makes AI methods even more attractive. In this study two AI tools, artificial neural networks (ANN) and genetic algorithms (GA), are utilized to develop models to predict bridge sufficiency ratings using current geometrical, age, traffic, and structural attributes as explanatory variables. Data is acquired from California Department of Transportation through the Internet and it includes 19120 structural bridge components owned and maintained by the State of California. The models developed by both ANN and GA provided promising and interpretable results. ANN models performed better when different models are constructed for different levels of sufficiency ratings. GA models outperformed ANN models while achieving a better goodness of fit even when using the whole data. However, remarkably prolonged training times for GA models might be considered as the only disadvantage for this type of application.