Comparison of decision learning models using the generalization criterion method.

Area of Science:

• Cognitive Science
• Decision Science
• Computational Neuroscience

Background:

• Accurate a priori predictions to novel conditions are essential for robust scientific models.
• Generalizability of decision learning models remains a key challenge in understanding human behavior.
• Evaluating model performance across different tasks and prediction horizons is crucial.

Purpose of the Study:

• To compare the generalizability of 8 decision learning models.
• To assess model performance in predicting behavior on unseen tasks.
• To determine if different models are optimal for short-term versus long-term predictions.

Main Methods:

• Participants completed two distinct gambling tasks: the Iowa Gambling Task and the Soochow Gambling Task.
• Models were evaluated using a priori predictions, with parameters estimated from one task applied to the other.
• Model performance was assessed using post hoc fit and generalization criteria for both short-term and long-term predictions.

Main Results:

• Models incorporating a prospect utility function exhibited significant generalizability to new conditions.
• Model performance varied depending on the prediction horizon (short-term vs. long-term).
• Distinct learning models were found to be more effective for short-term versus long-term predictions.

Conclusions:

• Decision learning models with prospect utility are capable of making generalizable predictions.
• The choice of learning model should consider the prediction timeframe (short-term vs. long-term).
• This research provides insights into model selection for decision-making tasks.