PUB - Publikationen an der Universität Bielefeld

Theory of Mind (ToM) enables people to infer other agents’ mental states such as their likely desire, intentions, beliefs and even emotions from observable behavior. Inferring such mental states is very important for social interactions, both in collaborative and competitive settings as it allows one to better adapt to other agents even without explicit communication. However, the act of inferring these mental states, also called mentalizing, is highly complex since any number of possible mental states could explain certain behavior equally well. Nevertheless, humans’ general ability to perform complex mental state inference has been shown in a range of studies. At the same time other studies have found that people exhibit cognitive biases, such as an egocentric bias in ToM tasks, especially in situations that require more intuitive or unsolicited ToM from participants. So far it is not entirely clear what causes this divergence between people’s general ToM capabilities and what they exhibit intuitively. In order to improve human-machine interactions, there exist a strong interest in equipping artificial systems with ToM capabilities. However, any such system needs to overcome the difficulties inherent in mentalizing to enable powerful but real-time capable ToM abilities suitable for interactions. Furthermore, such systems should be able to understand and adapt to people exhibiting cognitive biases. This thesis considers the problem of satisficing, i.e., satisfying and sufficing, mentalizing models for artificial systems. We argue that people’s mentalizing is based on general resource-sensitive cognition and that people thus are likely to (unconsciously) select different mental processes to perform mental state inference. We propose a similar adaptive process for mentalizing in artificial systems that can trade off resource requirements such as the available time with the accuracy of the mental state inference depending on the situation. In the first part of this thesis, we study cognitive biases in human mentalizing with a focus on implicit and unsolicited mentalizing. We find evidence that people appear to switch between different methods they employ for mental state inferences based on different factors which can result in cognitive biases. Additional results indicate that the type of mental state inference (explicit vs implicit) as well as the observed agent’s decision problem are potential factors that affect people’s mentalizing and their tendency to exhibit an egocentric bias. In the second part of this thesis, we build upon these findings to present computational models for satisficing ToM in artificial systems that can interpret human behavior efficiently under limited resources. The proposed models implement the idea of an adaptive system that dynamically selects which mentalizing model to use for a particular situation. We present different strategies of switching between suitable models for observational mentalizing and evaluate them on the empirical data from the first part. Our results indicate that such adaptive mentalizing can not only greatly reduce the computational costs, but also yield better accuracy in predicting observed behavior. Furthermore, in order to harness ToM for interactions, we present ideas about integrating satisficing mental state inferences with decision-making in a real-time-sensitive collaboration scenario. We evaluate a minimal implementation by integrating inferred mental states into an existing implicit behavior planning framework based on active inference.