Activity Sequence Modelling with Deep Generative Models

Abstract

Activity-based models typically structure the scheduling of activities as several sequential choices, e.g. number of activities, primary activity location and duration, secondary activities, etc. Conventional models cannot consider these choices simultaneously - limiting realistic interactions between them. Instead, we apply deep generative models for modelling activity sequence choices simultaneously, allowing full interaction between all dimensions. We evaluate two different data and model structures. The first uses an image-like representation of activity sequences, the second a text-like representation. We present results demonstrating the practical considerations as well as quality of these models in application. We use a Variational Auto-encoder architecture to provide realistic aggregate as well as dis-aggregate distributions. Our approach provides an alternative to discrete choice and scheduling based approaches for some applications. Our work also provides behavioural insight into the scheduling process.