Towards a Novel Classifier for the Representation of Bounded Rationality in Models of Travel Demand
Davy Janssens and Geert Wets
Several activity-based transportation models are now becoming operational and are entering the stage of application for the modelling of travel demand. In our application, we will use decision rules to support the decision-making of the model instead of principles of utility maximization, which means our work can be interpreted as an application of the concept of bounded rationality in the transportation domain. In this chapter we explored a novel idea of combining decision trees and Bayesian networks to improve decision-making in order to maintain the potential advantages of both techniques. The results of this study suggest that integrated Bayesian networks and decision trees can be used for modelling the different choice facets of a travel demand model with better predictive power than CHAID decision trees. Another conclusion is that there are initial indications that the new way of integrating decision trees and Bayesian networks has produced a decision tree that is structurally more stable.
Keywords: Bayesian networks; decision trees; BNT classifier; (un)supervised learning; rule complexity reduction
Introduction to Decision Theory
In decision theory in general, the predominant paradigm is expected utility theory (EUT) founded in von-Neumann and Morgenstern's utility theorem (McFadden, 2001). Here, a decision is considered to be a choice out of certain options, depending on the probability of occurrence and a valuation of alternatives. Proponents of the EUT approach praise its univocal theoretical foundation and its concomitant, clear mathematical interpretation enabling advanced statistical elaborations. However, even at the source of this theory, that is in economics, critics adhering to behavioural perspectives argue that although the rigid homo economicus assumption is useful in normative or prescriptive applications (as a model of how people ought to choose), people's everyday decision-making does not meet perfect rationality (Camerer, 1998). Indeed, the adoption of an EUT model implies a considerate, fully informed decision-maker, prone to a high degree of rationality; as opposed to approaches that account for bounded rationality (Simon, 1990), intuition (Plessner, Betsch, & Betsch, 2008) or uncertainty and lack of information of the decision-maker (Frederick, 2002; Tversky & Kahneman, 2002).
This dichotomy (rational versus behavioural) in theoretical approaches of decisionmaking applies to the different types of decisions that characterize individual travel as well. On the one hand is the repetitive nature of trips (such as commuting, chauffeuring kids to school, grocery shopping) likely to render (once) conscious decisions into script-based or habitual behaviour (Garling & Axhausen, 2003). On the other hand is activity scheduling (including choices of destinations, travel modes and routes) likely to entail the coordination of competing goals and intentions (e.g. amongst household members) in a complex environment (e.g. traffic-jams, opening hours), similar to complex planning problems (Garling, Gillholm, Romanus, & Selart, 1997). A theoretical account of behavioural decision-making in travel behaviour is given in Svenson (1998). The actual decision-making mechanisms of daily activity and travel scheduling and execution are scrutinized further in this book chapter.
To this end, we will demonstrate in this chapter a novel idea of combining decision trees and Bayesian networks to improve decision-making in travel behaviour and illustrate its application in The Albatross model (see Arentze & Timmermans, 2004) which is a rule-based model of activity-scheduling behaviour. The developed methodology can be seen as an example of bounded rationality in the sense that we assume with this approach that the decision-making of individuals is constrained by personal and scheduling obligations (e.g. bring/get children to school) as it is represented in the rule-based model of activity-scheduling behaviour that is used in our experiments.
The remainder of this chapter is organised as follows. First we will introduce the concept of activity-based (AB) models, which is followed by a discussion about how decision theory fits/is used within these models. Next, we will move to a section where Bayesian networks are introduced. In addition to the general concepts of the technique; more technical algorithms related to parameter and structural learning as well as entering evidences will be illustrated by means of concrete examples. In the fourth section, we will identify a problem of inferring decision rules from a Bayesian network, which is a typical problem in Bayesian network classifiers. In order to solve the problem, we will develop a novel integrated classifier in the fifth section of this chapter, in which we propose the idea to derive a decision tree from a Bayesian network (that is built upon the original data) instead of immediately deriving the tree from the original data. The next section describes both the data and the design of the experiments, which is then followed by an empirical (results) section. The chapter concludes with a discussion of the results.