Fast and Accurate Learning When Making Discrete Numerical Estimates.

Area of Science:

• Cognitive Science
• Psychology
• Decision Theory

Background:

• Everyday estimation tasks often involve discrete numerical quantities, such as counting objects or estimating discretized continuous variables.
• While Bayesian inference is common for continuous estimation, discrete numerical estimates are less studied despite their prevalence.

Purpose of the Study:

• To characterize how individuals learn discrete numerical distributions and make estimates using Bayesian decision theory.
• To investigate decision functions for converting uncertain representations into discrete numerical responses.

Main Methods:

• Employed Bayesian decision theory to model human estimation behavior.
• Utilized two tasks: a numerosity task and an area estimation task.
• Conducted three experiments with novel discrete stimulus distributions.

Main Results:

• Participants' response strategies aligned with either sampling from their posterior distribution or taking its maximum.
• Learned prior distributions were highly adaptive, with participants quickly acquiring discrete bimodal and quadrimodal priors.
• This learning occurred much faster than the thousands of trials typically needed for bimodal priors in continuous estimation tasks.

Conclusions:

• Discrete numerical estimation tasks serve as effective models for studying human learning and estimation processes.
• The rapid adaptation of prior distributions in discrete tasks offers new insights into cognitive learning mechanisms.