Humans depart from optimal computational models of socially interactive decision-making under partial information

Abstract

Decision making under uncertainty and under incomplete evidence in multiagent settings is of increasing interest in decision science, assistive robotics, and machine assisted cognition. The degree to which human agents depart from computationally optimal solutions in socially interactive settings is generally unknown. Yet, this knowledge is critical for advances in these areas. Such understanding also provides insight into how competition and cooperation affect human interaction and the underlying contributions of Theory of Mind. In this paper, we adapt the well-known ‘Tiger Problem’ from artificial-agent research to human participants in single agent and interactive, dyadic settings under both competition and cooperation. A novel element of the adaptation required participants to predict the actions of their dyadic partners in the interactive Tiger Tasks, to facilitate explicit Theory of Mind processing. Compared to computationally optimal solutions, participants gathered less information before outcome-related decision when competing with others and collected more evidence when cooperating with others. These departures from optimality were not haphazard but showed evidence of improved performance through learning across sessions. Costly errors resulted under conditions of competition, yielding both lower rates of rewarding actions and lower accuracy in predicting the actions of others, compared to prediction accuracy in cooperation. Taken together, the experiments and collected data provide a novel approach and insights into studying human social interaction and human-machine interaction when shared information is partial.

Learning to weigh competing moral motivations.

“…Applying formal learning frameworks to complicated social behaviors, including how we learn to trust or when we empathize, has been beneficial for gaining mathematical tractability in a fuzzy and complex psychological space (Cushman & Gershman, 2019). Indeed, the most striking examples of how formal learning models map onto sophisticated moral beliefs occur by carefully crafting learning algorithms and applying them to moral content (FeldmanHall & Nassar, 2021;Griffiths et al, 2010;Park et al, 2019;Steixner-Kumar et al, 2020;van Baar et al, 2022), ideally in ways that shape realworld issues, such as political polarization (Rathje et al, 2021). The success of such models in capturing learning as it relates to moral phenomena highlights the need for researchers across disciplines to embrace the role of social learning in shaping moral beliefs and the role of reward learning algorithms in reinforcing increasingly extreme moral judgments.…”