Identifying expectations about the strength of causal relationships

Area of Science:

• Cognitive Science
• Psychology
• Artificial Intelligence

Background:

• Bayesian models of causal induction depend on prior beliefs about causal systems.
• Understanding expectations of causal strength is crucial for these models.
• Previous proposals for prior distributions are difficult to test exhaustively.

Purpose of the Study:

• To estimate participants' prior beliefs about causal strengths using iterated learning.
• To establish a benchmark of human judgments for evaluating causal models.
• To compare the predictive accuracy of different Bayesian models.

Main Methods:

• Experiment 1: Iterated learning to estimate prior probability distributions of causal strengths.
• Experiment 2: Large-scale collection of human judgments on causal relationship strengths.
• Experiment 3: Estimation of prior beliefs across diverse causal systems.

Main Results:

• Iterated learning yielded prior distributions different from previous proposals.
• A Bayesian model using priors from iterated learning outperformed other models.
• Consistent prior belief patterns were observed across various causal scenarios.

Conclusions:

• Iterated learning provides a viable method for estimating prior beliefs in causal induction.
• The findings challenge existing assumptions about prior distributions in Bayesian causal models.
• Human causal expectations show systematic similarities across different contexts.