Modeling sensory-motor decisions in natural behavior

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Neuroscience

Background:

• Standard reinforcement learning models struggle with complex human behaviors due to environmental richness and cognitive limits.
• Human natural behaviors involve dynamic decision-making influenced by subjective values and sensory-motor processes.

Purpose of the Study:

• To propose a modular reinforcement learning model for more practical human behavior modeling.
• To develop a modular inverse reinforcement learning algorithm for estimating rewards and discount factors.
• To predict human navigation behaviors in virtual reality with high accuracy.

Main Methods:

• Developed a modular reinforcement learning framework.
• Created a modular inverse reinforcement learning algorithm to infer reward and discount parameters from human behavioral data.
• Validated the model's predictive accuracy across diverse subjects and tasks in virtual reality navigation.

Main Results:

• The modular model accurately predicts complex human navigation trajectories in novel environments.
• High accuracy was achieved across different subjects and tasks, demonstrating model generalizability.
• The model successfully reproduces intricate navigation paths using an artificial agent.

Conclusions:

• The proposed modular reinforcement learning model offers a more practical approach to modeling human reward-seeking and natural behaviors.
• This framework provides a strategy for estimating subjective action values and their impact on sensory-motor decisions.
• The findings have implications for developing more sophisticated artificial agents and understanding human decision-making processes.