Model-based algorithms shape automatic evaluative processing.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Reinforcement Learning

Background:

• Reinforcement learning algorithms are categorized as model-based (accurate, expensive) and model-free (cheap, error-prone).
• Prevailing theories link model-based algorithms to deliberate control and model-free to automatic control.
• This framework has influenced research across various psychological domains.

Purpose of the Study:

• To investigate the alignment between model-based/model-free algorithms and deliberate/automatic evaluative processing.
• To challenge the prevailing assumption that model-based algorithms exclusively underpin deliberate responses.

Main Methods:

• Three preregistered behavioral experiments with 2,572 adult participants.
• Replication of past findings and introduction of new experiments to control for confounds.
• Multinomial processing tree modeling to analyze automatic evaluation measures.

Main Results:

• Experiment 1 replicated previous findings but revealed confounding factors.
• Experiments 2 and 3, after eliminating confounds, demonstrated significant model-based contributions to automatic evaluations.
• Model-based algorithms were shown to shape both deliberate and automatic evaluative responses.

Conclusions:

• Dominant frameworks may underestimate the prevalence of model-based algorithms in decision-making.
• Automatic evaluative processing appears more computationally sophisticated than previously theorized.
• The distinction between automatic and deliberate systems may not strictly map to model-free and model-based algorithms.