Tobias Fabian Niehues¹ (), Dominik Straub^1,2, Constantin Rothkopf¹; ¹Technical University of Darmstadt, ²University of Cambridge

Human visually guided decision-making is influenced both by the desirability and the uncertainties of possible outcomes. Bayesian models of perception and action very often represent these quantities as quadratic costs and Gaussian prior beliefs about stimulus distributions. By contrast, several previous studies have suggested different non-quadratic functional forms for, e.g., sensorimotor learning and visual working memory tasks, as well as asymmetrical cost functions for incorporating resource-rational behavior, thereby aiming to explain often observed systematic biases. Incorporating more intricate cost functions makes it challenging to solve the decision-making problem and, consequently, renders it infeasible to solve the inverse problem of determining the parameters that govern subjects’ behavior, which include perceptual uncertainty, motor variability, cost functions, and prior beliefs about the stimulus. Here, we employ a recently introduced method that reverse-engineers various families of cost functions directly from observed behavioral data. This reconciles normative and empirical models of sensorimotor behavior. We analyze a variety of previously published experimental datasets and find that none of the observed behaviors is best described by the commonly assumed quadratic costs, but always by asymmetrical cost functions, most of which contain explicit costs representing effort costs for performing the movement. In most tasks, we found that prior beliefs are well-calibrated to the actual stimulus distributions and that subjects usually relied more on sensory information than on prior beliefs, but the degree to which they do and the variability across subjects within a task vary across tasks. Furthermore, we find that identifiability problems between priors and cost functions incorporating effort costs can be resolved by using behavioral data containing various levels of perceptual uncertainty, and show how our modeling framework can guide the design of experiments to best disambiguate alternative hypotheses about people’s implicit cost functions in visually guided behavior.

Acknowledgements: This research was supported by the European Research Council (ERC; Consolidator Award "ACTOR"-project number ERC-CoG- 101045783). We gratefully acknowledge the computing time provided to us on the high-performance computer Lichtenberg at the NHR Centers NHR4CES at TU Darmstadt.