Decoding behavior with minimal and interpretable agent models

Area of Science:

• Computational neuroscience
• Behavioral biology
• Systems biology

Background:

• Decision-making is a fundamental biological process across all scales.
• Understanding the internal mechanisms of decision-making is challenging.
• Existing methods often require detailed knowledge of internal states or environments.

Purpose of the Study:

• To develop a general method for reconstructing decision-making processes from behavioral data alone.
• To create an agent model based on recurrent dynamics over internal states.
• To apply this method to diverse biological agents without prior internal mechanism knowledge.

Main Methods:

• Defining agent models with recurrent dynamics over discrete internal states.
• Encoding and processing sensory information within these states.
• Dictating actions based on internal state computations.
• Validating the method on synthetic agents and experimental data.

Main Results:

• The method accurately reconstructs agent behavior from observations.
• Successfully inferred agent models for rats (evidence accumulation) and mice (decision-making under uncertainty).
• Few internal states were sufficient to accurately reproduce observed behaviors.

Conclusions:

• The developed method offers a powerful tool for understanding decision-making across biology.
• It provides interpretable insights into the computational processes underlying behavior.
• Applicable to diverse organisms, advancing comparative studies of cognition.