hide
Free keywords:
-
Abstract:
Meta-cognition, i.e., our ability to assess the quality of our own decisions, is a critical contributor to the regulation of choice. Various ways of formalizing the sensitivity and bias of our confidence judgements have duly been related to neural processing and have helped distinguish cognitive disorders. However, most of these studies focus on immediate rather than sequential choice – for instance, perceptual decision-making or value-based decision-making for known outcomes. Here, we treat sensitivity and bias for value-based decision-making problems in which outcome values must be learned across trials. We repurpose the central idea underlying the meta-cognitive assessment measure, meta-d’. Crudely, d’ quantifies the sensitivity of choice (of an ‘actor’) in perceptual decision-making; and meta-d’ quantifies the sensitivity of confidence judgements (made by a ‘rater’), by measuring how well the rater could have made the original choice (interpreting confidence as a form of probability judgement). In the learning case, we build two computational models: a Forward model, characterizing the subjects’ choices and generating ‘first order’ confidence from the modelled probability of being correct; and a Backward model, which generates choices whose first-order confidence best matches the subjects’ confidence reports. The Performance of Backward and Forward models are in the roles of meta-d’ and d’ in our measure of meta-cognitive sensitivity, called MetaRL.Ratio. Our results demonstrated that MetaRL.Ratio was consistence with previous measures of meta-cognitive sensitivity and also differentiated high and low meta-cognitive behavior. This study suggests that MetaRL.Ratio is a promising tool for assessing meta-cognitive sensitivity in the value-based learning domain.