Modular inverse reinforcement learning for visuomotor behavior.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Robotics

Background:

• Modeling animal and human behavior requires biologically grounded computational approaches.
• Reward-based learning, particularly in visuomotor tasks, is a key mechanism for goal-directed behavior.
• Existing models often lack biological plausibility, necessitating methods rooted in empirical evidence.

Purpose of the Study:

• To develop a computational model for behavior that integrates biological principles with learning mechanisms.
• To quantify implicit goals in observed behavior using inverse reinforcement learning.
• To propose a modular inverse reinforcement learning algorithm for dissecting complex behaviors like navigation.

Main Methods:

• Utilizing a reinforcement learning framework to model goal-directed behavior.
• Applying inverse reinforcement learning to infer reward functions from observed trajectories.
• Developing a modular inverse reinforcement learning algorithm to estimate reward contributions from component tasks (path following, obstacle avoidance, target approach).

Main Results:

• Successfully recovered component reward weights for individual navigation tasks.
• Demonstrated that behavioral goals can succinctly explain variability in observed trajectories.
• Showed that the model can achieve good estimates with limited observational data.

Conclusions:

• The proposed modular inverse reinforcement learning approach provides an expressive and accurate model of behavior rooted in biological learning principles.
• This method allows for the prediction of behavior in novel situations based on inferred reward structures.
• The findings support a modular cognitive architecture and highlight the power of inverse reinforcement learning in understanding goal-directed actions.