Designing optimal behavioral experiments using machine learning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Decision Science

Background:

• Computational models offer precise theories of human cognition but challenge traditional experimental design.
• Existing methods struggle to effectively test and compare complex computational models.

Purpose of the Study:

• To provide a tutorial on using Bayesian optimal experimental design (BOED) and machine learning for designing informative experiments.
• To demonstrate a method for efficiently evaluating computational models and their parameters against data.

Main Methods:

• Leveraging advances in BOED and machine learning to find optimal experimental designs for any model generating simulation data.
• Applying the approach to study exploration-exploitation trade-offs in multi-armed bandit tasks.

Main Results:

• Optimal designs derived via BOED more efficiently distinguish between competing models of human behavior.
• The method efficiently characterizes behavior given a specific model, outperforming common experimental designs.
• Simulations and a real-world experiment validated the BOED approach.

Conclusions:

• BOED provides a powerful framework for advancing computational modeling in cognitive science.
• The presented methodology enhances model comparison and parameter estimation efficiency.
• Practical guidance and code are provided for broader adoption.