Randomization and Adaptive Learning in a Simplified Poker Game

Abstract

Thirty pairs of subjects participated in three different two-person zerosum Poker games in extensive form with imperfect and asymmetric information. The results provide no support for the (unique) mixed-strategy equilibrium solution for risk-neutral players on either the individual or the aggregate level. Compared to this solution, the informed players do not bluff as often as they should, and the uninformed players call too often. Comparison of the present study with previous studies of diametrically opposed conflicts shows consistent differences between the extensive form of two-person zerosum games and its strategically equivalent normal form. An adaptive learning model proposed by Roth and Erev (1995) succeeds in tracking the general trends in the probabilities of bluffing and calling over time and in accounting for the effects of size of stake, when it assumes that learning takes place only on stage games when the players are called upon to act.