Probabilistic Learning in a Volatile Environment in Individuals with and without Autism Spectrum Disorder

Background: Predictive coding has recently been proposed as a framework to understand Autism Spectrum Disorder (ASD). It assumes that the human mind processes information by making and testing predictions, and that violations to these predictions result in prediction errors. One possible explanatory account argues that individuals with ASD tend to ascribe High and Inflexible Precisions to Prediction Errors (HIPPEA), irrespective of the context. Ascribing high weights or precisions to all prediction errors could be an advantageous strategy for probabilistic learning within a stable environment, but would hamper learning in a volatile and unstable environment.

Objectives: In order to gain insight into these mechanisms of precision-setting in an unstable environment in individuals with ASD, we administered a Probabilistic Learning Task in a Volatile Environment (PLTVE). Based on the Probabilistic Reversal Learning task of Behrens et al. (2007), the PLTVE is designed to evaluate whether ASD individuals can estimate the volatility of the context and use this estimate to flexibly adjust their precision-setting of prediction errors. To investigate the influence of noise on precision-setting in ASD, we extended this task by adding variability in irrelevant dimensions of the presented stimuli. Based on the HIPPEA framework, we hypothesize differences in probabilistic learning to be more pronounced when variability in irrelevant dimensions is present.

Methods: In a behavioral experiment we presented the PLTVE to an ASD group (11-27 y) and a typically developing (TD) group (13-26 y). The PLTVE is a two-alternative forced choice task, in which the probabilistic rule about which feature out of three potentially relevant dimensions is rewarded, changes unannounced. The number of (ir)relevant dimensions and whether the relevant dimension was cued or not was manipulated in three experimental phases (uncued unidimensional - cued three-dimensional - uncued three-dimensional).

Results: Overall, no differences in consistency on the PLTVE task were found between the TD and the ASD group over all 3 phases. In phase 1 and phase 2, the average consistency of both groups approximated the reward probabilities. In phase 3, average consistency was lower than the reward probability for both groups. However, results revealed an interaction effect of group and lose-shift/win-stay strategy, with TD participants making more use of the win-stay strategy and ASD participants making more use of the lose-shift strategy. In addition, participants in the ASD group were more likely to give up trying to find the correct rule in phase 3, compared to participants in the TD group.

Conclusions: The present results suggest that individuals with and without ASD can adapt to a more volatile environment. When variability was increased, both TD and ASD individuals had more difficulties with probabilistic learning. Interestingly, our results reveal that individuals with ASD make less use of positive feedback (i.e. prediction confirmation) and more use of negative feedback (prediction error) in their choice for the next trial, indicating that they do assign different weights to most recent trials. In addition, individuals with ASD seem to withdraw more frequently when confronted with high amounts of uncertainty (phase 3).