Active inference and the two-step task.

Area of Science:

• Cognitive Science
• Neuroscience
• Computational Psychiatry

Background:

• Sequential decision-making involves learning and balancing exploration-exploitation.
• Reinforcement learning is a common model, but may not fully capture human behavior in tasks like the two-step task.
• Active inference offers an alternative framework for decision-making under uncertainty.

Purpose of the Study:

• To investigate if active inference provides a more accurate model of human behavior in the two-step task compared to reinforcement learning.
• To re-analyze existing datasets using Bayesian model selection.

Main Methods:

• Re-analysis of four publicly available two-step task datasets.
• Bayesian model selection to compare active inference and reinforcement learning models.
• Assessment of behavioral model predictions against empirical data.

Main Results:

• Active inference better described two datasets exhibiting model-based inference and directed exploration.
• Learning via probability distributions contributed to improved model fits.
• Approximately 50% of participants demonstrated sensitivity to information gain under active inference.

Conclusions:

• Active inference offers a potentially more accurate computational model for certain aspects of human behavior in sequential decision-making.
• The findings support the empirical validation of active inference.
• Further research into alternative models for the two-step task is warranted.