Novel Neural Network Applications to Mode Choice in Transportation: Estimating Value of Travel Time and Modelling Psycho-Attitudinal Factors

Whenever researchers wish to study the behaviour of individuals choosing among a set of alternatives, they usually rely on models based on the random utility theory, which postulates that the single individuals modify their behaviour so that they can maximise of their utility. These models, often identified as discrete choice models (DCMs), usually require the definition of the utilities for each alternative, by first identifying the variables influencing the decisions. Traditionally, DCMs focused on observable variables and treated users as optimizing tools with predetermined needs. However, such an approach is in contrast with the results from studies in social sciences which show that choice behaviour can be influenced by psychological factors such as attitudes and preferences. Recently there have been formulations of DCMs which include latent constructs for capturing the impact of subjective factors. These are called hybrid choice models or integrated choice and latent variable models (ICLV). However, DCMs are not exempt from issues, like, the fact that researchers have to choose the variables to include and their relations to define the utilities. This is probably one of the reasons which has recently lead to an influx of numerous studies using machine learning (ML) methods to study mode choice, in which researchers tried to find alternative methods to analyse travellers’ choice behaviour. A ML algorithm is any generic method that uses the data itself to understand and build a model, improving its performance the more it is allowed to learn. This means they do not require any a priori input or hypotheses on the structure and nature of the relationships between the several variables used as its inputs. ML models are usually considered black-box methods, but whenever researchers felt the need for interpretability of ML results, they tried to find alternative ways to use ML methods, like building them by using some a priori knowledge to induce specific constrains. Some researchers also transformed the outputs of ML algorithms so that they could be interpreted from an economic point of view, or built hybrid ML-DCM models. The object of this thesis is that of investigating the benefits and the disadvantages deriving from adopting either DCMs or ML methods to study the phenomenon of mode choice in transportation. The strongest feature of DCMs is the fact that they produce very precise and descriptive results, allowing for a thorough interpretation of their outputs. On the other hand, ML models offer a substantial benefit by being truly data-driven methods and thus learning most relations from the data itself. As a first contribution, we tested an alternative method for calculating the value of travel time (VTT) through the results of ML algorithms. VTT is a very informative parameter to consider, since the time consumed by individuals whenever they need to travel normally represents an undesirable factor, thus they are usually willing to exchange their money to reduce travel times. The method proposed is independent from the mode-choice functions, so it can be applied to econometric models and ML methods equally, if they allow the estimation of individual level probabilities. Another contribution of this thesis is a neural network (NN) for the estimation of choice models with latent variables as an alternative to DCMs. This issue arose from wanting to include in ML models not only level of service variables of the alternatives, and socio-economic attributes of the individuals, but also psycho-attitudinal indicators, to better describe the influence of psychological factors on choice behaviour. The results were estimated by using two different datasets. Since NN results are dependent on the values of their hyper-parameters and on their initialization, several NNs were estimated by using different hyper-parameters to find the optimal values, which were used to verify the stability of the results with different initializations.