Parameters, Predictions, and Evidence in Computational Modeling: A Statistical View Informed by ACT-R

Area of Science:

• Computational Cognitive Psychology
• Cognitive Modeling
• Statistical Validation

Background:

• Current model validation in cognitive psychology often uses methods from experimental physics.
• This approach can obscure participant and theory variability, creating a "model fixed, data variable" paradigm.
• This makes interpreting predictions and individual differences challenging.

Purpose of the Study:

• To propose a novel likelihood-based, "data fixed, model variable" paradigm for computational model validation.
• To address limitations in current validation methods that obscure sources of variation.
• To provide a framework for integrating statistical and cognitive modeling approaches.

Main Methods:

• Treating computational models as stochastic processes.
• Accounting for participant-to-participant variation within experiments.
• Applying a likelihood-based framework to model validation.

Main Results:

• The proposed "data fixed, model variable" paradigm offers a more robust approach to model validation.
• This framework can be applied to diverse mechanistic cognitive architectures.
• Demonstrated with a focus on a class of ACT-R models.

Conclusions:

• The new paradigm facilitates better interpretation of model predictions and individual differences.
• It encourages a shift towards stochastic validation in cognitive modeling.
• Aims to foster communication between statisticians and cognitive modelers.