Decision Theory in AB Models
Decision theory in AB travel demand modelling can be represented by three major lines of research as described in Rasouli and Timmermans (2014). The first line of research is represented by constraints-based models. Constraints-based models do not predict individual activity-travel patterns but check whether any given activity agenda is feasible to meet specific (space—time) constraints. These models are not very strong from a behavioural point of view and the earliest examples of these models go back to PESASP (Lenntorp, 1976); CARLA (Jones, Dix, Clarke, & Heggie, 1983) and BSP (Huigen, 1986). In recent years several contributions by researchers active in space—time geography have led to interesting contributions (see e.g. Neutens, Schwanen, Witlox, & De Maeyer, 2008; Soo, Zhang, Ottens, & Ettema, 2009) but these contributions are net yet integrated in full operational AB models. Secondly, the main research line of the representation of decision theory in AB models is represented by utility-maximizing models. These models are based on the premise that individual maximize utility in choosing between activity-travel pattern alternatives. Significant contributions in the field of AB travel demand modelling are, for example, the daily activity schedule model (Ben-Akiva, Bowman, & Gopinath, 1996); PCATS (Kitamura & Fujii, 1998) and CEMDEP (Bhat, Guo, Srinivasan, & Sivakumar, 2004), FAMOS (Pendyala, Kitamura, Kikuchi, Yamamoto, & Fujii, 2005) and SIMAGENT (Goulias et al., 2012). Finally, computational process models (CP Ms) try to overcome the drawback of utility-based models, namely that travellers do not make 'optimal' decisions but rather context- dependent heuristic decisions. CPMs '... replace the utility-maximizing framework with behavioural principles of information acquisition, information representation, information processing, and decision making' (Golledge, Kwan, & Garling, 1994). CPMs are basically also microsimulations due to their disaggregate nature, the sequential decision process and the use of heuristics. However, the heuristics employed by CPMs rather consist of 'if-then' rules than utility-maximizing decision criteria. Models in this line of research are SCHEDULER (Golledge et al., 1994), AMOS (Kitamura & Fujii, 1998; Pendyala, Kitamura, Reddy, & Chen, 1995), ALBATROSS (Arentze & Timmermans, 2004), FEATHERS (Bellemans, Janssens, Wets, Arentze, & Timmermans, 2010), Tasha (Miller & Roorda, 2003) and ADAPTS (Auld & Mohammadian, 2009).
Within the context of CPMs, and more concretely within the Albatross and Feathers ABM's, a series of Ph.D. dissertations and studies (Janssens, 2005; Moons, 2005; Sammour, 2013) have been conducted in the field of machine learning. The machine learning technique in question then typically produces an individual decision-making heuristic/rule set as output and as such using a typical <if-then- else > -structure, is a nice representation of the concept of bounded rationality.
Bounded rationality goes back to the work of Simon (1991), where the concept was introduced as having two interlocking components: the limitations of the human mind and the structure of the environments in which the mind operates. The first component means that models of human judgment and decision-making should be built on what we actually know about the mind's capacities. In many real-world situations, like for instance in daily travel decisions, optimal strategies are unknown or unknowable (Simon, 1987). Because of the mind's limitations, humans 'must use approximate methods to handle most tasks' (Simon, 1990, p. 6). These methods include recognition processes that largely obviate the need for further information search, heuristics that guide search and determine when it should end, and simple decision rules that make use of the information found. Machine learning techniques are very well suited to perform these tasks. The second component of Simon's view of bounded rationality, environmental structure, is of crucial importance because it can explain when and why simple heuristics perform well: if the structure of the heuristic is adapted to that of the environment. Since we agree that there is no such thing as a uniform law of travel behaviour which explains daily activity-travel patterns all over the world; this is exactly what we do when we model travel behaviour. By incorporating a wide spectrum of geographical, contextual and land use characteristics in our models, our goal is to adapt the structure of the heuristic to that of the environment. The next section focuses on one particular approach (Bayesian networks) within the field of machine learning that is able to represent this concept of bounded rationality.