Task-set Selection in Probabilistic Environments: a Model of Task-set Inference

To act effectively in a complicated, uncertain world, people often rely on task-sets (TSs) that define action policies over a range of stimuli. Effectively selecting amongst TSs requires assessing their individual utility given the current world state. However, the world state is, in general, latent, stochastic, and time-varying, making TS selection a difficult inference for the agent. An open question is how observable environmental factors influence an actor's assessment of the world state and thus the selection of TSs. In this work, we designed a novel task in which probabilistic cues predict one of two TSs on a trial-by-trial basis. With this task, we investigate how people integrate multiple sources of probabilistic information in the service of TS selection. We show that when action feedback is unavailable, TS selection can be modeled as “biased Bayesian inference”, such that individuals participants differentially weight immediate cues over TS priors when inferring the latent world state. Additionally, using the model’s trial-bytrial posteriors over TSs, we calculate a measure of decision confidence and show that it inversely relates to reaction times. This work supports the hierarchical organization of decision-making by demonstrating that probabilistic evidence can be integrated in the service of higher-order decisions over TSs, subsequently simplifying lower-order action selection.