Improving the Performance of a Multi-Agent Rule-Based Model for Activity Pattern Decisions Using Bayesian Networks

Several activity-based models are now becoming operational and are entering the stage of application in transport planning. Some of these models use a set of decision trees to support its decision making instead of using principles of utility maximization. However, it is believed that the structure of decision trees can sometimes be very instable and sensitive of highly correlated predictors. Therefore, in this study it is examined whether decision trees constitute the best representational form to capture the behavioural mechanisms and principles that individuals and households use to organize their activities. The paper reports the findings of experiments that were conducted by means of Bayesian networks to gain a better understanding in the predictive performance of Albatross, which is a sequential rule based model of activity scheduling behaviour. The performances of Bayesian networks and decision trees are compared and results are evaluated by means of a detailed quantitative and qualitative analyses. The results have shown that Bayesian networks outperformed the decision tree based approach for all decision agents of the Albatross model. Given the excellent performance, we believe that the research community may potentially consider the use of Bayesian networks in developing existing or future activity-based transportation models. Janssens, Wets, Brijs, Vanhoof, Arentze and Timmermans 3