Travel Mode Choice Modeling Based on Improved Probabilistic Neural Network

Abstract

Travel mode choice can be regarded essentially as a problem of pattern recognition. In the past, numerous researchers applied artificial neural network (ANN) to travel mode choice modelling. Feed-forward back propagation neural network (BPNN) is widely used, and most of the studies show that BPNN presents better prediction accuracy and generalization capacity. Probabilistic neural network (PNN) is also known as a general solution to pattern classification problems by following an approach developed in statistics. Comparing to BPNN, its shorter training time and better stability increase the reliability of simulation results, even so PNN is barely used in travel behaviour analysis. Therefore the paper tries to apply PNN for travel mode choice modelling. First, network structure is established based on the data obtained from resident trip survey, and then the K-means cluster algorithm is applied to optimize the hidden node number so that they can be modified dynamically. Moreover, the data centers and extended constants of gaussian radial basis function are changed adaptively during the learning progress, and the improved PNN is called KPNN. To reveal the superiority of KPNN, BPNN is also introduced to make a comparison. For BPNN, the optimal number of hidden nodes is recognized, and then the well trained network is applied. By proposing two performance measures, the classification and predictive capability of KPNN and BPNN is evaluated, it is found that KPNN outperforms BPNN on both the simplicity and prediction accuracy. Anyway, the simulation results show that KPNN can deal with the problem of travel choice modelling as well as, if not better than BPNN.