Integrating rough sets with neural networks for weighting road safety performance indicators

Abstract

This paper aims at improving two main uncertain factors in neural networks training in developing a composite road safety performance indicator. These factors are the initial value of network weights and the iteration time. More specially, rough sets theory is applied for rule induction and feature selection in decision situations, and the concepts of reduct and core are utilized to generate decision rules from the data to guide the self-training of neural networks. By means of simulation, optimal weights are assigned to seven indicators in a road safety data set for 21 European countries. Countries are ranked in terms of their composite indicator score. A comparison study shows the feasibility of this hybrid framework for road safety performance indicators.