Experience-Weighted Attraction Learning in Coordination Games: Probability Rules, Heterogeneity, and Time-Variation

Area of Science:

• Behavioral Economics
• Game Theory
• Econometrics

Background:

• Previous research introduced the experience-weighted attraction (EWA) learning model for predicting behavior in economic experiments.
• The EWA model demonstrated superior fit compared to existing learning models like choice reinforcement and belief-based models.
• Prior estimations assumed stationary learning parameters and a representative agent approach.

Purpose of the Study:

• To enhance the EWA learning model by incorporating nonstationary learning parameters.
• To investigate player heterogeneity by allowing for distinct parameter values across player segments.
• To compare the performance of logit and power probability response functions in predicting choices.

Main Methods:

• Econometric estimation of the EWA model using experimental data from weak-link and median-action coordination games.
• Relaxation of the stationary parameter assumption to allow for time-varying learning parameters.
• Comparison of logit (exponential) and power probability response functions.

Main Results:

• Player behavior exhibits heterogeneity, with distinct learning parameter segments observed.
• Learning parameters were found to adjust only slightly over time.
• Logit probability response functions consistently performed as well as or better than power functions.

Conclusions:

• The refined EWA learning model, accounting for heterogeneity and nonstationary parameters, provides a more accurate prediction of dynamic behavior in economic games.
• The logit response function is a more effective tool for transforming strategy attractions into choice probabilities.
• Findings suggest that player learning dynamics are complex and exhibit subtle temporal adjustments.