Assisted specification of discrete choice models with piece-wise linear utility functions

Abstract

This paper extends RUMBoost, a model that enhances discrete choice models (DCMs) by automatically learning non-linear parameters using gradient boosting, to piece-wise linear functions. The model output is used to specify DCMs with piece-wise linear utility functions. The methodology is demonstrated using the Swissmetro dataset, showing significant differences in behaviour over models with linear utility function specifications.