An information-theoretic account of human–computer interaction

This thesis presents a theoretical framework for the study of interactive systems, using methods from information theory, machine learning and control theory. The framework builds on the information-theoretic capacities of empowerment, relevant information and mutual information, which I adapt and apply to the domain of human-computer interaction. Three user studies exploring dynamic interactive scenarios - one car-tracking and two collaborative target-acquisition experiments - provide empirical data for the development of probabilistic models, used in the characterisation of specific aspects of human performance, such as the level of control, the quality of decision-making, and the level of engagement in interpersonal coordination. Human control models are extended to accommodate for the inherent lags, characteristic for human-computer and human-human interaction, in a principled way. Optimal controllers, describing particular patterns of human behaviour, are built on these theoretical models, providing evidence for specific limits of human performance through simulations. The thesis describes the potential of empowerment, as a generic task-independent measure of control, to characterise the uncertainty in human-machine interfaces. This work builds an important bridge between theory and experiments, and suggests that the proposed information-theoretic concepts could provide analytical tools for supporting the design and evaluation of interactive systems, by elucidating novel aspects of human performance complementing standard measures. The thesis provides proof of concept examples for the application of such information-theoretic measures, and demonstrates how they can be treated naturally side-by-side along traditional metrics used in HCI research. It emphasises the acquisition cost of accurate theoretical models, necessary to ensure the reliability of such measures