Data-integrated simulation of interactive behaviour

Computational modelling of interactive behaviour represents a key challenge in human-computer interaction (HCI) but dominant models – whether predictive, based on cognitive science, or datadriven – fail to fully account for the complex dynamics of interactive behaviour, require timeconsuming manual adaptation of model parameters or large amounts of training data. A promising alternative is the framework of computational rationality that formalises behaviour as a sequential decision making process of a rational agent learning to interact autonomously constrained by a set of cognitive, environmental, or behavioural bounds. First works have successfully demonstrated the potential of this approach for HCI but only for well-defined interactive tasks in carefully controlled settings. The goal of this project is to advance these simulator models and make them applicable to more complex, everyday interactive tasks by merging cognitive models with data-driven methods. Our starting point is to formulate interactive behaviour as partially observable Markov decision processes (POMDP) of self-learning agents. Taking inspiration from humans, we plan to solve these using a combination of model-based and model-free approaches. We will also investigate methods to infer cognitive model parameters from human behaviour as well as to explain the underlying reasoning for agent decisions.