The interpretation of computational model parameters depends on the context.

Area of Science:

• Cognitive and brain sciences
• Computational neuroscience
• Developmental psychology

Background:

• Reinforcement learning (RL) models are widely used in cognitive and brain sciences to explain behavior and cognition.
• Contradictory findings in the RL literature raise questions about the validity of current assumptions.
• Common assumptions include parameter generalizability across contexts and interpretability of neurocognitive processes.

Purpose of the Study:

• To investigate the generalizability and interpretability of computational model parameters in reinforcement learning.
• To test whether RL parameters generalize across different learning tasks and capture distinct neurocognitive processes.
• To examine the developmental trajectories of RL parameters in young individuals.

Main Methods:

• Recruited 291 participants aged 8-30 years.
• Administered three distinct learning tasks within a single experimental session.
• Fitted reinforcement learning models to individual participant data for each task.

Main Results:

• Some parameters, like exploration/decision noise, showed significant generalization and similar developmental trajectories across tasks.
• Generalization levels were notably below the methodological maximum, indicating limitations.
• Other parameters, including learning rates and forgetting, lacked generalization and sometimes exhibited opposing developmental trends.
• Low interpretability was observed for all tested RL parameters.

Conclusions:

• The generalizability and interpretability of computational cognitive models are limited by current assumptions about parameters.
• Contextual factors, such as reward stochasticity and task volatility, significantly influence parameter behavior.
• Systematic investigation of context factors is crucial for advancing the reliability and explanatory power of RL models in cognitive science.