Epiphany Learning for Bayesian Updating: Overcoming the Generalized Monty Hall Problem

Abstract

We create a modified version of the Monty Hall problem, which separates the “Irrelevant, therefore invariant” heuristic from true Bayesian updating (while the original one does not), and conduct laboratory experiments to show how a 100-door variant of the problem helps people learn to play optimally (always switch). Experimental results show that after playing the 100-door variant, subjects obtain an average switching rate above 80% in the 3-door problem, higher than most of the previous work without subject communication and/or competition. Moreover, results from estimating structural learning models using subject-level data show that the individual learning process is more likely to be an epiphany rather than a gradual one, such as reinforcement learning.