Parallel model-based and model-free reinforcement learning for card sorting performance.

Area of Science:

• Cognitive Neuroscience
• Computational Psychiatry

Background:

• The Wisconsin Card Sorting Test (WCST) is a key measure of cognitive flexibility.
• Perseveration errors on the WCST, repeating a sorting rule after negative feedback, are linked to reduced cognitive flexibility.
• Previous research indicates response demands can modulate the frequency of perseveration errors.

Purpose of the Study:

• To propose and evaluate parallel reinforcement-learning models for WCST performance.
• To investigate how model-free (response-level) and model-based (category-level) learning interact.
• To compare these new models against existing attentional-updating and purely model-based approaches.

Main Methods:

• Development of parallel reinforcement-learning models incorporating both response-level and category-level learning.
• Comparison of model predictive accuracy using data from 375 participants on a computerized WCST.
• Analysis of how models account for the modulation of perseveration by response demands.

Main Results:

• Parallel reinforcement-learning models demonstrated superior predictive accuracy for most participants.
• These models uniquely explained the observed modulation of perseveration propensity by response demands.
• Existing models, including attentional-updating, did not fully capture these dynamics.

Conclusions:

• Parallel reinforcement-learning models offer a novel theoretical framework for understanding WCST performance.
• This framework provides a more nuanced view of cognitive flexibility and behavioral adaptation.
• The models facilitate the identification of individual differences in the underlying latent processes governing flexibility.