One and done? Optimal decisions from very few samples.

Area of Science:

• Cognitive Science
• Decision Theory
• Computational Neuroscience

Background:

• Human behavior in learning and inference tasks often aligns with Bayesian ideal observer models.
• A discrepancy exists where humans appear to use limited samples from probability distributions, rather than the full distribution, for decision-making.
• This limited sampling seems insufficient for accurate probability distribution approximation under standard Bayesian assumptions.

Purpose of the Study:

• To investigate the optimal number of samples humans should use when making decisions based on costly sampling within statistical decision theory.
• To reconcile observed human sampling-based behavior with the Bayesian inference hypothesis of cognition.
• To explore the trade-offs between decision speed, sample cost, and overall reward optimization.

Main Methods:

• Analysis within the framework of statistical decision theory.
• Modeling decision-making under resource constraints (time costs of sampling).
• Evaluating strategies for optimizing total expected or worst-case reward over numerous decisions.

Main Results:

• Under assumptions of time costs for sampling, a strategy of making numerous rapid decisions using very few samples can be globally optimal.
• This approach prioritizes frequent, albeit locally suboptimal, decisions over fewer, more computationally intensive, optimal decisions.
• The findings suggest that frequent, low-information decisions can maximize long-term reward.

Conclusions:

• The study reconciles observed human sampling behaviors (e.g., probability matching) with Bayesian cognitive models.
• Resource-constrained cognition, where limited samples are used for efficiency, provides a viable explanation for suboptimal-looking decision-making.
• Future research should focus on resource-constrained cognitive processes and their impact on decision-making strategies.