Friday, 5th April 2019
Opponent's Foresight and Optimal Choices
Using two experiments, we demonstrate that in a “winner-take-all” extensive-form game, where players’ payoffs are directly opposed, optimal behavior can deviate from the unique “sure-win” backward-induction strategy due to beliefs about the opponent’s expertise. In the game we use, a particular deviation from the “sure-win” strategy yields a higher payoff only if one’s opponent makes a mistake; otherwise, the deviation leads to a loss. We find that experienced subjects were more likely to deviate when they were informed that their opponent was inexperienced and, in a different experiment, when they inferred their opponent’s inexperience by only observing their opponent’s preceding moves in the game. Maximum likelihood estimation indicates that Rampal’s (2018) model of limited foresight and uncertainty about the opponent’s foresight fits the data better than the Dynamic Level-k (Ho and Su (2013)) and Agent Quantal Response Equilibrium (McKelvey and Palfrey (1998)) models.