טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentAvivit Bercovici-Boden
SubjectLearning in Ensemble of Decision Problems
DepartmentDepartment of Industrial Engineering and Management
Supervisors Full Professor Tennenholtz Moshe
Full Professor Erev Ido
Full Thesis textFull thesis text - English Version


Abstract

We propose a machine learning approach to action prediction in one-shot games. In contrast to the huge literature on learning in games where an agent's model is deduced from its previous actions in a multi-stage game, we propose the idea of inferring correlations between agents' actions in different one-shot games in order to predict an agent's action in a game which she did not play yet. We introduce a comparative study of several   methods for action prediction in \emph{one-shot} games. We provide a    igorous framework where prediction methods can be defined and evaluated. We consider two leading research lines in experimental economics and cognitive psychology, which model how a player chooses a strategy (the Cognitive hierarchy model and the model proposed by Costa-Gomes et al.), a prediction rule which is based on populations statistics, and a prediction method based on a machine learning   approach. The machine learning approach which we introduce proposes the idea of inferring correlations between agents' actions in different one-shot games. We show, using three data sets obtained in experiments with human subjects (two of them presented in [9] and [15] and one experiment that we conducted), the advantages of the latter method. Furthermore, we demonstrate that the learning method can be used to increase payoffs of an adequately informed agent. Our comparative study makes use of the data gathered by the founders of the different approaches, and compare the different approaches on all of these data sets, with four different evaluation criteria, offering a complete rigorous comparison. This is, to the best of our knowledge, the first comparison of such methods, and the first application of learning methods for obtaining high payoffs in one shot games.