On Predictive Planning and Counterfactual Learning in Active Inference.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Cognitive Science

Background:

• Understanding intelligent behavior is crucial with rapid AI advancements.
• Active inference provides a theoretical framework for sophisticated planning and decision-making.
• Existing models often focus on either planning or learning from experience.

Purpose of the Study:

• To investigate two decision-making schemes within active inference: planning and learning.
• To introduce a novel mixed model combining planning and learning for balanced decision-making.
• To evaluate the model's adaptability in a challenging grid-world scenario.

Main Methods:

• Examined two distinct decision-making strategies in active inference.
• Developed a mixed model integrating planning and learning.
• Evaluated model performance in a grid-world task requiring agent adaptability.
• Analyzed parameter evolution for insights into decision-making processes.

Main Results:

• The proposed mixed model demonstrates balanced decision-making by integrating planning and learning.
• The model shows adaptability in a challenging grid-world environment.
• Analysis of parameter evolution offers insights into the decision-making framework.

Conclusions:

• The mixed active inference model offers a principled and adaptable approach to intelligent decision-making.
• This framework contributes to explainable AI by providing insights into decision-making processes.
• The study highlights the benefits of combining planning and learning for robust behavior.