A Computational Framework for Modeling Belief-based Decision Making

Thesis (Ph.D.)--University of Washington, 2021Existing computational models of decision making are often limited to particular experimental setups. The limitation is mainly due to the inability to capture the decision maker's uncertainty about the situation. We propose a computational framework for studying decision making under uncertainty in neuroscience and psychology. Our framework is heavily focused on the probabilistic assessment of the decision maker, i.e., their "belief", about the state of the world. Specifically, it is based on Partially Observable Markov Decision Processes (POMDPs), which combines Bayesian reasoning and reward maximization to choose actions. We demonstrate the viability of our belief-based decision making framework using data from various experiments in perceptual and social decision making. Our framework explains the relationship between decision makers' actual performance and their belief about it, called decision confidence, in perceptual decision making experiments. It also shows why this assessment could deviate from reality in many situations. Such deviations have been often interpreted as evidence for sub-optimal decision making or distinct processes that underlie choice and confidence. Our framework challenges these interpretations by showing that a normative Bayesian decision maker optimizing the gained reward elicits the same discrepancies. Moreover, our method outperforms existing models in quantitatively predicting human behavior in a social decision making task and provides insight into the underlying process. Our results suggest that in decision making tasks involving large groups, humans employ Bayesian inference to model the "group's mind" and make predictions of others' decisions. Finally, we extend our method to multiple reasoning levels about others (levels of theory of mind) and make the connection to conformity as a strategy for decision making in groups. This extended framework can explain human actions in various collective group decision making tasks, providing a new theory for cooperation and coordination in large groups