Modeling Complex Animal Behavior with Latent State Inverse Reinforcement Learning.

Area of Science:

• Neuroscience
• Computational Biology
• Animal Behavior

Background:

• Understanding complex animal behavior is key to linking brain activity to actions.
• Current models struggle with long-term, naturalistic behaviors like navigation.
• A need exists for dynamic modeling of extended animal movements.

Purpose of the Study:

• To introduce discrete Dynamical Inverse Reinforcement Learning (dDIRL), a novel paradigm for modeling long-term animal behavior.
• To apply dDIRL to understand navigation in water-starved mice.
• To infer internal states and reward functions driving observed behaviors.

Main Methods:

• Developed dDIRL, a latent state-dependent framework for behavior modeling.
• Utilized expectation-maximization to infer reward functions and state transitions from observed data.
• Applied the model to individual water-starved mice navigating a labyrinth.

Main Results:

• Identified three distinct internal states governing mouse behavior.
• Found a consistent water-seeking state, though not always dominant.
• Discovered two distinct clusters of animals based on labyrinth exploration patterns.

Conclusions:

• dDIRL provides a nuanced understanding of how internal states and rewards shape complex behaviors.
• The model offers insights into the neural basis of naturalistic navigation.
• This framework advances the study of dynamic animal behavior in intricate environments.