MODE CHOICE ANALYSIS USING NEURAL NETWORK

Abstract

The analysis of mode choice for urban commuters is useful in accurately predicting the share of trips carried by each mode, so that planning efficient of transportation system can be facilitated. The study of choice behaviour of commuters is an important stage in transportation planning and quite complex due to the inherent complexities of human behavour. Though it is impossible to understand or predict totally the hum-an behaviour, yet information regarding important factors that influence behaviour towards travel related options can be obtained using theories of consumer behaviour. Generally the evaluation of travel behaviour of commuters is carried out by using disaggregate Multinomial Logit Model (MNL). This traditional mode choice model is complex and circuitous due to non-linear dependence relationship. The predictive ability of this models goes down as the number of alternatives increases. The motivation for adding neural network at a new modelling methodology stems from its apparent relevance to problems requiring large scale, highly dimensional data analysis, such as travel related behaviour. The approach of neural network is non parameteric, and do not require any assumptions about the functional form of the underlying distribution of data. Therefore tools like Artificial Neural Network modelling becomes more attractive. This dissertation demonstrates the use of Artificial Neural Network for mode choice analysis. For this purpose, BPN programme have been written by developed using object-oriented programming language (C++). By using this programme four back propagation models were developed to understand the travel behaviour of commuter of various vehicle ownership groups. These models were trained and tested on travel behaviour data,of Delhi commuters, available at Centre of Transportation Engineering, (COTE) University of Roorkee, Roorkee. The black box image of Neural Network is iv somewhat clarified by finding out the relative importance of various input variables used in the present study through partitioning of weights algorithm. The same data sets which were used to develop ANN model, have also been used in calibrating and validating the of MNL models. The comparative study has been made on the following approaches. Comparison between ANN and MNL analysis Comparison between models developed by using BPN program written for present study purpose and Neural planner 4.1.package. On the basis of the results obtained in this study, ANN based forecasting models, were found superior to traditional MNL models