Discipline hop for high-fidelity, high-generalisation models of human behaviour

Mathematical models of human behaviour are crucial in areas like transport, economics, robotics, and epidemiology to predict the impact of new technologies or policies.

These models are typically machine-learned from large datasets or mechanistic, based on assumptions about human cognition. However, machine-learned models require extensive real-world data, and cognitive models, while effective in specific scenarios, do not scale well to diverse real-world situations. Neither approach can generalise to novel situations, such as interactions with new technologies.

Recent advances in cognitive and machine-learned modelling suggest a best-of-both-worlds approach that can achieve both high-fidelity behaviour emulation and strong generalisation to new situations. This approach views human behaviour as boundedly optimal, maximising rewards within perceptual, motor, and cognitive constraints.

By integrating mechanistic models from computational cognitive science and deep reinforcement learning methods, this new research program aims to model complex real-world behaviour. The project, led by a PI with expertise in road traffic safety and vehicle automation, will be hosted at King’s College London within a team specialising in ICT domains.

The primary objective is to establish a cross-disciplinary bridge for the envisioned research program. A proof of concept will be developed, focusing on safety-relevant pedestrian-vehicle interaction using data from Leeds City Council. The project aims to enhance human behaviour modelling capabilities in the UK, with significant academic and societal impact.