Latent goal models for dynamic strategic interaction.

Area of Science:

• Decision neuroscience
• Computational neuroscience
• Animal behavior

Background:

• Traditional studies often use simplified tasks with discrete choices and known optimal solutions.
• Naturalistic interactions involve continuous actions, dynamic environments, and unknown optimal strategies.

Purpose of the Study:

• To bridge the gap between simplified experimental paradigms and complex, naturalistic agent interactions.
• To develop and validate a novel computational model for analyzing social dynamics in continuous action spaces.

Main Methods:

• Developed a "penalty shot" video game task for pairs of monkeys competing in real-time.
• Modeled monkey strategies using a goal-based dynamical system with stochastically evolving goals.
• Employed approximate Bayesian inference and neural networks to parameterize player goals.

Main Results:

• The model successfully captured complex player dynamics and generated realistic synthetic data.
• Strategy complexity, measured by change points, varied more across sessions than within sessions.
• More complex offensive strategies were beneficial, while defensive strategies did not show similar benefits.

Conclusions:

• The developed experimental paradigm and computational model provide a powerful toolkit for studying realistic social dynamics.
• This approach allows for the investigation of decision-making in more naturalistic and complex interactive environments.
• Findings offer insights into how strategy complexity influences performance in competitive social interactions.