Feature-based Choice and Similarity in Normal-form Games: An Experimental Study.

In this paper we test the effect of descriptive "features" on initial strategic behavior in normal form games, where "descriptive" are all those features that can be modified without altering the (Nash) equilibrium structure of a game. We observe that our experimental subjects behave according to some simple heuristics based on descriptive features, and that these heuristics are stable even across strategically different games. This suggests that a categorization of games based on features may be more accurate in predicting agents' initial behavior than the standard categorization based on Nash equilibria, as shown by the analysis of individual behavior. Analysis of choice patterns and individual response times suggests that non-equilibrium choices may be due to the use of incorrect and simplified mental representations of the game structure, rather than to beliefs in other players' irrationality. Of the four stationary concepts analyzed (Nash equilibrium, QRE, action sampling, and payoff sampling), QRE results the best in fitting the data.