PUB - Publikationen an der Universität Bielefeld

Theory of Mind is the cognitive ability to reason about the mental state of others. It allows us to infer not only others' intentions but also their (potentially false) beliefs about us or parts of our environment. Developing artificial systems with similar capabilities is of great interest both from a cognitive science perspective, as artificial models can be used to test hypotheses about this mentalizing capability of humans, as well as a human-machine-interaction perspective: Inferring a user's intention without the need for explicit communication will improve ease of use and collaboration. Inferring even more information about mental states which may hold a potential false belief, allows the system to provide more precise help or information. One popular framework for designing such artificial systems capable of reasoning about mental states is the Bayesian Theory of Mind by (Baker et al., 2017). This framework has been successfully employed in a range of different scenarios by using specifically designed mental models. So far these scenarios have usually been fairly simple as Bayesian inference becomes intractable quickly as the number of considered mental states increases. Therefore, one currently faces two problems when trying to expand the framework to broader usecases: 1. Designing mental models applicable to a wide range of scenarios and 2. finding ways to quickly perform inferences in these mental models in order to allow for their use in real-time interaction systems. In the present work, we are trying to find solutions to both of these problems: In previous work (Pöppel & Kopp 2018) we have shown that a system capable of switching between simple models, each corresponding to specific assumptions about another agent's mental states, can explain human navigation data efficiently. We are currently evaluating sampling approaches in more complex and general models to achieve similar results. By sampling discrete mental states for each of the considered belief states inference becomes feasible again. We can use the likelihood of the observed data given the sampled mental states to decide whether to stick to the current sample or modify it, effectively employing a "Win stay, loose switch" strategy which has previously been found in human causal learning (Levine, 1975). Building upon this sampling approach we are evaluating if the way the sample changes to fit the observed behavior allows us to infer previously unknown mental states in order to allow the system to adapt to changes in the scenario.