Adaptive and Satisficing Cognition for Theory of Mind in Interaction

Pöppel J, Kopp S (2021)
In: Program of the Computational Cognition Workshop 2021 (ComCo 2021).

Kurzbeitrag Konferenz / Poster | Englisch
 
Download
Es wurden keine Dateien hochgeladen. Nur Publikationsnachweis!
Abstract / Bemerkung
When agents interact with each other, they often depend on some form of "understanding'' of each other. For competitive interactions, insight into the intentions of the other agent can give one an edge. For cooperative settings, understanding the intentions and capabilities of team members is important to maintain coordination. In either case an agent needs to make suitable decisions based on beliefs about others' mental states. The concept of inferring others' mental states is a cognitive capacity that has been termed Theory of Mind (ToM). People develop increasingly sophisticated mentalizing capabilities from inferring others' likely goals to their (potentially false) beliefs, preferences and even emotions (Wellman & Liu 2004) which can even be considered recursively. Nevertheless, it appears as if people do not always invest all the computational costs involved in these complex inferences. For example, evidence has suggested that people often employ heuristics, e.g., in the form of egocentric biases, in their decision making when not explicitly prompted to consider others' mental state. In our own work, we have found that people appear to "switch'' between different modes of ToM depending on a wide range of factors, including priming a mental state consideration (Pöppel & Kopp 2019) as well as the observed agent's decision problem (Pöppel et al. 2021). This view of mentalizing is in line with a resource-rational (Lieder & Griffiths 2019) perspective on human reasoning. Exact inferences of complex mental states are computationally very demanding (Blokpoel et al. 2013). At the same time, the inferred knowledge may not always be relevant for a given situation or the potential costs of making incorrect inferences may be negligible. In that sense people may be performing "satisficing'' mentalizing where they try to only invest as much mental effort as required for the current situation. For artificial agents to be able to interact with others in socially intelligent ways, they also need to be equipped with those capabilities. The Bayesian Theory of Mind (BToM) framework already provides promising results when it comes to explaining the inferential capabilities of people (Baker et al. 2017). However, BToM is inherently computationally very demanding and relies on strong assumptions and heuristics to be applicable (Baker et al. 2009). These computational costs are more severe in interactions where people expect timely reactions of others. Further, in scenarios with independent agents, long computation times may lead to reasoning about outdated states and thus unsuitable actions. In (Pöppel et al. submitted) we have shown how a minimal BToM model coupled with a predictive procession hierarchy can enable low-level coordination with real-time capabilities in a situated interaction task. For these reasons we present first steps towards a computational approach to adaptive and satisficing ToM. Such models can, on the one hand, better account for how humans mentalize about others ecologically and, on the other hand, enable artificial agents to efficiently draw the other-related inferences needed in competitive or cooperative interaction. In earlier work, we successfully explored the use of a "switching'' approach that employs different inference models that make different assumptions (or heuristics) (Pöppel & Kopp 2018). The agent sticks with simpler models as long as they are capable of explaining its observations, and only switches to more complex ones when needed. This switching model was able to outperform any single specialized model as well as a full BToM model that was not restricted to specific assumptions while being orders of magnitude more efficient than the full model. However, the model repertoire was limited to a specific scenario and the switching strategy rather simplistic. Work is underway to improve on these strategies in an actual interaction setting, where an agent may also want to adapt its strategies for planning as well as mentalizing based on what the interaction partners do.
Stichworte
Theory of Mind; Adaptive Reasoning; Satisficing Mentalizing
Erscheinungsjahr
2021
Titel des Konferenzbandes
Program of the Computational Cognition Workshop 2021 (ComCo 2021)
Konferenz
Computational Cognition 2021
Konferenzort
Osnabrück, Germany
Konferenzdatum
23.09.2021 – 24.09.2021
Page URI
https://pub.uni-bielefeld.de/record/2957587

Zitieren

Pöppel J, Kopp S. Adaptive and Satisficing Cognition for Theory of Mind in Interaction. In: Program of the Computational Cognition Workshop 2021 (ComCo 2021). 2021.
Pöppel, J., & Kopp, S. (2021). Adaptive and Satisficing Cognition for Theory of Mind in Interaction. Program of the Computational Cognition Workshop 2021 (ComCo 2021)
Pöppel, J., and Kopp, S. (2021). “Adaptive and Satisficing Cognition for Theory of Mind in Interaction” in Program of the Computational Cognition Workshop 2021 (ComCo 2021).
Pöppel, J., & Kopp, S., 2021. Adaptive and Satisficing Cognition for Theory of Mind in Interaction. In Program of the Computational Cognition Workshop 2021 (ComCo 2021).
J. Pöppel and S. Kopp, “Adaptive and Satisficing Cognition for Theory of Mind in Interaction”, Program of the Computational Cognition Workshop 2021 (ComCo 2021), 2021.
Pöppel, J., Kopp, S.: Adaptive and Satisficing Cognition for Theory of Mind in Interaction. Program of the Computational Cognition Workshop 2021 (ComCo 2021). (2021).
Pöppel, Jan, and Kopp, Stefan. “Adaptive and Satisficing Cognition for Theory of Mind in Interaction”. Program of the Computational Cognition Workshop 2021 (ComCo 2021). 2021.

Export

Markieren/ Markierung löschen
Markierte Publikationen

Open Data PUB

Suchen in

Google Scholar