Bayesian generic priors for causal learning.

Area of Science:

• Cognitive Science
• Psychology
• Artificial Intelligence

Background:

• Human causal learning involves inferring cause-effect relationships from observations.
• Existing Bayesian models often lack explicit assumptions about the structure of causal systems.
• Prior research suggests humans favor simpler explanations for observed phenomena.

Purpose of the Study:

• To propose and evaluate a Bayesian model of causal learning incorporating generic priors.
• To investigate how these priors influence judgments of causal strength and structure.
• To compare the proposed model against alternative Bayesian and non-normative models.

Main Methods:

• Developed a Bayesian model with generic priors favoring sparse and strong (SS) causes.
• Utilized a causal generating function assuming independent unobservable causal influences.
• Tested the model by fitting it to multiple experimental datasets with varied parameters.

Main Results:

• The SS power model successfully accounted for human judgments of both causal strength and causal structure.
• The model explained why causal power and effect base rates influence structure judgments more than sample size.
• Compared to a Bayesian model without generic priors, the SS model showed improved data fit.

Conclusions:

• Generic priors favoring sparse and strong causes are crucial for accurate Bayesian causal learning models.
• The proposed SS power model provides a robust framework for understanding human causal inference.
• The findings have broader implications for understanding human learning within a Bayesian framework.