Forecasting travel demand: A comparison of logit and artificial neural network methods

This paper describes the use of backpropagation artificial neural networks to forecast travel demand from disaggregate discrete choice data and compares them with logit models. Three data sets are used; synthetic data which fulfil the underlying logit assumptions, synthetic data which breach the underlying logit assumptions and real data. It is found that neural networks with no hidden layers exhibit almost identical performance to logit models in all three cases. For the synthetic data which breach the underlying logit assumptions and with real data, backpropagation neural networks with a hidden layer can achieve a better fit than logit. However, careful choice of the number of hidden units and training iterations is needed to avoid overfitting and consequent degradation of performance.