Toward inverse generative social science using multi-objective genetic programming.

Area of Science:

• Computational Social Science
• Agent-Based Modeling
• Behavioral Economics

Background:

• Generative mechanism-based models, including agent-based simulations, necessitate specifying intra-agent equations.
• Existing methods often present numerous equation choices within a mechanism class, impacting model interpretability.
• Balancing empirical accuracy with theoretical enlightenment is crucial for robust generative models.

Purpose of the Study:

• To develop and demonstrate a novel method for simultaneously exploring empirical fit and theoretical interpretability in agent-based models.
• To automate the discovery of intra-agent equations that optimize both predictive power and theoretical clarity.
• To investigate the trade-off between these objectives in the context of social norms and alcohol use behavior.

Main Methods:

• Implementation of a multi-objective genetic programming approach.
• Application to an existing agent-based simulation of alcohol use behaviors informed by social norms theory.
• Evolution of intra-agent equations to simultaneously optimize for empirical prediction and theoretical interpretability.

Main Results:

• Successful demonstration of the multi-objective genetic programming method in evolving agent-based model equations.
• Discovery of a discernible trade-off between the empirical accuracy (fit) and theoretical interpretability of the evolved models.
• Identification of specific social norms processes influencing alcohol use dynamics, derived from the model's structure.

Conclusions:

• The developed method effectively automates the exploration of competing objectives in generative modeling.
• The identified trade-off provides valuable insights into the complexities of social norms influencing behavior change and stability.
• This approach enhances the theoretical enlightenment of agent-based simulations beyond mere historical replication.