Interpretation of Bayesian neural networks for predicting the duration of detected incidents

Abstract

ABSTRACTThis study introduces Bayesian learning to neural networks for accurate prediction of incident duration. Network parameters are updated using a hybrid Monte Carlo algorithm, and yield reasonable accuracy with mean absolute percentage error of 29%. A pedagogical rule extraction algorithm (TREPAN) is applied to extract comprehensible representations from the neural networks. The TREPAN facilitates better comprehensibility with M-of-N expression, and maintains high predictive accuracy to its respective network. Extracted decision trees provide a discovery and explanation of previously unknown relationships present in incident nature, and represent a series of decisions to assist traffic management operators in better decision making. Furthermore, to quantify the importance of variables from the neural network, a connection weight approach is used. Factors appearing in the first splitter of decision tree show high relative importance, indicating that they are influential for longer or shorter incident duration. Interpretation of Bayesian neural networks is an important addition to the Advanced Traveler Information Systems toolkit.