TARK Conference 2017 Conference Paper
- Sujata Ghosh
- Aviad Heifetz
- Rineke Verbrugge
- Harmen de Weerd
In an earlier experiment, participants played a perfect information game against a computer, which was programmed to deviate often from its backward induction strategy right at the beginning of the game. Participants knew that in each game, the computer was nevertheless optimizing against some belief about the participant's future strategy. In the aggregate, it appeared that participants applied forward induction. However, cardinal effects seemed to play a role as well: a number of participants might have been trying to maximize expected utility. In order to find out how people really reason in such a game, we designed centipede-like turn-taking games with new payoff structures in order to make such cardinal effects less likely. We ran a new experiment with 50 participants, based on marble drop visualizations of these revised payoff structures. After participants played 48 test games, we asked a number of questions to gauge the participants' reasoning about their own and the opponent's strategy at all decision nodes of a sample game. We also checked how the verbalized strategies fit to the actual choices they made at all their decision points in the 48 test games. Even though in the aggregate, participants in the new experiment still tend to slightly favor the forward induction choice at their first decision node, their verbalized strategies most often depend on their own attitudes towards risk and those they assign to the computer opponent, sometimes in addition to considerations about cooperativeness and competitiveness.