Computational cognitive modeling involves creating formal, mathematical, or computational models that simulate aspects of human cognition. By implementing theories as running programs or mathematical equations, modelers transform vague verbal theories into precise, testable predictions. Models must specify exactly how information is represented, processed, stored, and retrieved — forcing theoretical precision that verbal theories can avoid.
Modeling Frameworks
Production systems (ACT-R, Soar) model cognition as sequences of condition-action rules that operate on symbolic representations in working and long-term memory. Connectionist models (neural networks) represent knowledge as patterns of activation across interconnected processing units. Bayesian models frame cognition as rational statistical inference, explaining behavior as optimal or near-optimal given the learner's prior knowledge and available evidence. Drift-diffusion models capture the dynamics of decision-making as noisy evidence accumulation toward decision boundaries. Reinforcement learning models explain how organisms learn from reward and punishment.
A key challenge in computational modeling is model comparison: how do we determine which model best accounts for the data? Bayesian model comparison (using metrics like the Bayesian Information Criterion or marginal likelihood) penalizes model complexity to prevent overfitting. Parameter recovery analyses verify that model parameters can be uniquely estimated from data. Qualitative predictions (patterns of results that distinguish models) complement quantitative fit statistics.
Contributions
Computational models have advanced understanding across cognitive domains: drift-diffusion models reveal how speed-accuracy trade-offs arise from decision threshold adjustment; ACT-R models predict the time course of problem solving; Bayesian models explain how people combine prior knowledge with new evidence; and reinforcement learning models reveal how reward prediction errors drive learning. The growing integration of computational models with neural data creates increasingly constrained and testable theories.