When mice think twice: optimal control explains when and how mice overcome default biases

Abstract

Decisions permeate every aspect of our lives but decision-making policies seem to vary tremendously across time and individuals. Rather than processing all decision-relevant information, we sometimes rely on habitual and/or intuitive responses, which can lead to irrational biases and errors. This raises the twofold issue of whether a single normative principle can (i) determine when cognitive control is engaged, and (ii) account for the diverse behavioral signature of cognitive control regulation. In line with recent theoretical proposals, our working hypothesis is that cognitive control is flexibly deployed to optimize a cost-benefit trade-off. Using Markov Decision Processes, we derive the optimal online control allocation policy in standard decision-making tasks, under arbitrary default biases. This bridges the gap between motivational theories of cognitive control regulation and accumulation-to-bound models of decision making. Under the model, response times, default response rates, decision accuracy and interference effects are not independent behavioral phenomena, but distinct manifestations of the same underlying arbitration process. We then test the distinct behavioral predictions of the model using a customized pseudo-ecological set-up, whereby mice engage with a self-administered perceptual decision-making task over several weeks. In this context, processing decision-relevant cues engages cognitive control, whereas default responses are determined by learned, cue-independent, action-outcome associations. The model accurately predicts both performance and response time interference effects, their interactions, as well as their modulation by reward and task demands. This work provides novel computational means for assessing the brain circuits that operate the arbitration between default and controlled decision processes, in rodent models of both healthy and neuropsychiatric conditions.