Machine Learning in Transport Economics

Most transportation research techniques have been developed in the 1960s and 1970s, but our computational capabilities have vastly grown since then. A lot has happened in statistics and new techniques have been developed since then. Application of those new approaches in transport economics seems to be slow.

I have lately been learning different prediction techniques, which have advantages over more traditional approaches. Decision trees and neural networks are capable of capturing non-linear relationships in data, with much lower levels of error than traditional k-nearest neighbours, linear and logistic regression approaches.

Application of these techniques in transport economics has not been wide, but certainly has potential. Especially with more transportation data becoming available to researchers.

To me training of a decision tree or a neural network still seems to be a challenging art to me. Challenges come from selection of hyperparameters. Some approaches exist for optimisation of hyperparameters in machine learning, but it seems that the traditional approach of grid search, which is simply searching through manually selected subset of hyperparameter space for the parameters that return a model with the lowest error, seems to work ok for the models that I have been calculating lately.

I am looking forward to new projects in the future to apply the machine learning approaches that I have been learning.