Generative AI Agents for Travel Behaviour: Applications in Surveys and Modelling

Abstract

We explore the potential of Generative Artificial Intelligence (AI) agents created using open-access and locally hosted Large Language Models (LLMs) in replicating human survey behaviour and mode choice preferences in scenario-based travel surveys. The aim is to establish performance and validation benchmarks for utilizing AI agents in travel behaviour analysis, agent-based simulations, and other case uses. Accordingly, we developed a systematic scientific approach to assess the performance of seven open-access foundational LLMs, with parameters ranging from one to seventy billion, which can be generalized for creating and validating the performance of Generative AI agents in various applications. The AI agents were developed using a zero-shot learning approach, incorporating both unrestricted sociodemographic and static prompting, as well as a dynamic restricted sociodemographic prompting strategy. The performance of these agents was validated against the human benchmark dataset, evaluating their effectiveness and reliability in capturing and replicating nuanced travel behaviour.