Robustness of artificial neural network and discrete choice modelling in presence of unobserved variables

Morgane Dumont, Johan Barthelemy, Timoteo Carletti

Research output: Contribution in Book/Catalog/Report/Conference proceedingConference contribution

30 Downloads (Pure)

Abstract

Models are used to gain a better understanding of complex systems such as the evolution of a population, the transportation demand, the brain behaviour, elections outcome, the propagation of a disease,... System models should be precise and parsimonious. However, the total variation of the system cannot be precisely captured by the observed variables as there can be unobserved ones influencing the system output. The unexplained variation caused by unobserved variables is, therefore, considered as a noise in the model. Different models handle that noise in a different way. For instance, a linear regression assumes that the noise follows a normal distribution and explicitly incorporates it into the model formulation. On the other hand, other models, such as a deterministic neural network, do not explicitly incorporate that noise. Several models can then be applied and the selection of the best one can be a challenging question. This research aims to highlight the importance of the unobserved variables on the results of two types of simple yet widely used models: feedforward neural networks (FFNN) and logit discrete choice models (LDCM). The first application consists in modelling the divorces in an agent-based microsimulation, the agents being the individuals of a given population. For each couple in the model, the divorce is predicted based on the characteristics of the couple (ex: length of the marriage, age of the individuals). In this application, it is shown that the LDCM outperforms the neural network due to the presence of - possibly many - unobserved variables. The second example is a model defined to predict the level of interaction between groundwater and quarry extensions. In this application, the value of every relevant variable is assumed to be known, i.e. the noise from unobserved variables is minimum. In this case, it is shown that both approaches perform well, but FFNN perform slightly better than LDCM. We then investigate how the model performance evolves when the noise increases by removing variables from the models specification. Finally, those two applications will allow us to conclude on the robustness of the discrete choice models and artificial neural network in presence of unobserved variables.
Original languageEnglish
Title of host publicationProceedings - 22nd International Congress on Modelling and Simulation, MODSIM 2017
Subtitle of host publicationModelling and Simulation Society of Australia and New Zealand
EditorsGeoff Syme, Darla Hatton MacDonald, Beth Fulton, Julia Piantadosi
Pages480-486
Number of pages7
ISBN (Electronic)978-0-9872143-7-9
Publication statusPublished - 1 Jan 2017

Publication series

NameProceedings - 22nd International Congress on Modelling and Simulation, MODSIM 2017

Keywords

  • Discrete choice modelling
  • Neural network
  • unobserved variables
  • Unobserved variables

Fingerprint

Dive into the research topics of 'Robustness of artificial neural network and discrete choice modelling in presence of unobserved variables'. Together they form a unique fingerprint.

Cite this