Choice variability and suboptimality in uncertain environments

Making decisions under uncertainty, from perceptual judgments to reward-guided choices, requires combining multiple pieces of decision-relevant information — a cognitive process modeled as statistical inference. In such conditions, human and animal decisions exhibit a large suboptimal variability whose origin and structure remains poorly understood. This variability is usually hypothesized as noise at the periphery of inferential processes, namely sensory noise in perceptual tasks and stochastic exploration in reward-guided learning, or as suboptimal biases in inference per se. Here we outline a theoretical framework aiming at characterizing the origin and structure of choice variability in uncertain environments, with an emphasis on the computational imprecision of inferential processes usually overlooked in the literature. We indicate how to modify existing computational models and behavioral paradigms to dissociate computational imprecisions from suboptimal biases in inference. Computational imprecisions have critical consequences for understanding the notion of optimality in decision-making.