Heuristics as Bayesian inference under extreme priors.

Area of Science:

• Cognitive science
• Decision-making
• Machine learning

Background:

• Simple heuristics are often favored for their tractability, despite ignoring data.
• The "less-is-more" effect, where simpler models outperform complex ones, is common in cognitive science.
• It's debated whether ignoring information offers an inherent computational advantage.

Purpose of the Study:

• To investigate why simple heuristics can outperform complex models.
• To formally analyze the relationship between heuristics and Bayesian inference.
• To explore the optimal balance between information use and simplification in decision-making.

Main Methods:

• Formal Bayesian analysis to model decision strategies.
• Comparison of heuristics (tallying, take-the-best) with Bayesian inference and linear regression.
• Simulations across a continuum of Bayesian models with varying prior strengths.

Main Results:

• Heuristics are equivalent to Bayesian inference with infinitely strong priors.
• Intermediate Bayesian models, balancing information use and priors, outperformed both heuristics and full-information models in simulations.
• Discarding information is never computationally optimal.

Conclusions:

• Heuristics' success stems from their strong priors approximating environmental structure, not simplicity.
• Down-weighting information with appropriate priors is superior to ignoring it.
• New heuristics can be derived by strengthening priors in Bayesian models, with implications for psychology, machine learning, and economics.