Humans versus computer algorithms in repeated mixed strategy games

This paper is concerned with the modeling of strategic change in humans’ behavior when facing different types of opponents. In order to implement this efficiently a mixed experimental setup was used where subjects played a game with a unique mixed strategy Nash equilibrium for 100 rounds against 3 preprogrammed computer algorithms (CAs) designed to exploit different modes of play. In this context, substituting human opponents with computer algorithms designed to exploit commonly occurring human behavior increases the experimental control of the researcher allowing for more powerful statistical tests. The results indicate that subjects significantly change their behavior conditional on the type of CA opponent, exhibiting within-sub jects heterogeneity, but that there exists comparatively little between-subjects heterogeneity since players seemed to follow very similar strategies against each algorithm. Simple heuristics, such as win-stay/lose-shift, were found to model subjects and make out of sample predictions as well as, if not better than, more complicated models such as individually estimated EWA learning models which suffered from overfitting. Subjects modified their strategies in the direction of better response as calculated from CA simulations of various learning models, albeit not perfectly. Examples include the observation that subjects randomized more effectively as the pattern recognition depth of the CAs increased, and the drastic reduction in the use of the win-stay/lose-shift heuristic when facing a CA designed to exploit this behavior.