Calibrating a New Reinforcement Learning Mechanism for Modeling Dynamic Activity-Travel Behavior and Key Events

Abstract

Recent travel demand modeling mainly focuses on activity-based modeling. However the majority of such models are still quite static. Therefore, the current research aims at incorporating dynamic components, such as short-term adaptation and long-term learning, into these activity-based models. In particular, this paper attempts at simulating the learning process underlying the development of activitytravel patterns. Furthermore, this study explores the impact of key events on generation of daily schedules. The learning algorithm implemented in this paper uses a reinforcement learning technique, for which the foundations were provided in previous research. The goal of the present study is to release the predefined activity-travel sequence assumption of this previous research and to allow the algorithm to determine the activity-travel sequence autonomously. To this end, the decision concerning transport mode needs to be revised as well, as this aspect was previously also set within the fixed schedule. In order to generate feasible activity-travel patterns, another alteration consists of incorporating time constraints, for example opening hours of shops. In addition, a key event, in this case “obtaining a driving license”, is introduced into the learning methodology by changing the available set of transport modes. The resulting patterns reveal more variation in the selected activities and respect the imposed time constraints. Moreover, the observed dissimilarities between activity-travel schedules before and after the key event prove to be significant based on a sequence alignment distance measure.