The Effect of Uncertainty on Prediction Error in the Action-Perception Loop

Abstract

Abstract Among all their sensations, agents need to distinguish between those caused by themselves and those caused by external causes. The ability to infer agency is particularly challenging under conditions of uncertainty. Within the predictive processing framework, this should happen through active control of prediction error that closes the action-perception loop. Here we use a novel, temporally-sensitive, behavioural proxy for prediction error to show that it is minimised most quickly when variability is low, but also when volatility is high. Further, when human participants report agency, they show steeper prediction error minimisation. We demonstrate broad effects of uncertainty on accuracy of agency judgements, movement, policy selection, and hypothesis switching. Measuring autism traits, we find differences in policy selection, sensitivity to uncertainty and hypothesis switching despite no difference in overall accuracy.