Activity Based Modelling with Deep Conditional Generation

Abstract

Modelling human activity scheduling is a challenging task at the core of activity-based modelling. Existing approaches to activity scheduling are increasingly expensive and slow to develop, and can also produce unrealistically homogenous outputs, failing to model the real diversity in human behaviours. We contribute a novel methodology combining a deep generative model with conditionality, such that the model can be used in an activity modelling or transport simulation based framework. By explicitly and simultaneously modelling variation of observed activity schedules, we better represent real diversity. Our experimental results demonstrate that our approach is both cheaper and faster than existing activity scheduling solutions, whilst still providing closely tailored and high quality outputs.