Decision, Inference, and Information: Formal Equivalences Under Active Inference.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computational Neuroscience

Background:

• A core challenge is unifying principles for inference, learning, and action.
• Active inference proposes variational free energy minimization as a unifying principle.
• Existing literature on active inference's links to other fields is fragmented.

Purpose of the Study:

• To systematically review and develop conceptual links between active inference and other optimal behavior frameworks.
• To expose a shared optimization principle underlying decision-making and information processing.
• To orient researchers new to active inference and clarify foundational assumptions.

Main Methods:

• Systematic literature review and conceptual analysis.
• Formal correspondence analysis of major decision-making and information processing frameworks.
• Variational formulation of expected free energy minimization.

Main Results:

• Active inference formally subsumes classical models like Bayesian decision theory, optimal control, and reinforcement learning.
• Connections are established with information-theoretic principles such as rate-distortion theory and maximum entropy.
• Shared optimization principles are identified across diverse theoretical frameworks.

Conclusions:

• Active inference provides a unifying framework for understanding inference, learning, and action.
• The review synthesizes fragmented literature, facilitating cross-disciplinary integration.
• This work clarifies active inference's foundational assumptions and broad applicability.