טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentMergui Lahmi Meghan
SubjectPrediction of Individual Choice Behavior
DepartmentDepartment of Industrial Engineering and Management
Supervisors ASSOCIATE PROF. Tamir Hazan
DR. Ori Plonsky
Full Thesis textFull thesis text - English Version


Abstract

Understanding how humans make decisions, and being able to apply behavioral economic theory is a fundamental task in many domains. For behavioral economics to become a truly applied science, there is a need to predict people’s response when presented with some choice setting. In our research we are focusing on predicting individual human behavior when facing choice between gambles. Specifically, we focused on predicting choice behavior in simple abstract economic games. A recent prediction tournament, CPC18, described by Plonsky et al. challenged researchers on such games.

During our research we investigated several methods in order to predict individual human behavior on this set of games. We created a new set of individual behavior features, and tried machine learning models, but we were not able to outperform the baseline with this method. We classified games according to their specificities and trained different models for each subclass independently. With this method, we were able to slightly outperform the baseline on this prediction task. Another attempt was to create a synthetic dataset of games and individual answers. We used this synthetic dataset to model individual parameters based on games already played, and to train a prior model that retrieves the individual behaviors of players. We use the first model to assign individual parameters on the real players, and the prior model to give prediction on the outputs of new games. We were not able to assign enough behavioral parameters on the real players, and were not able to outperform the baseline on this task, using this method.

In our research we experimented with new complex approaches that were inspired and emerged from the work of excellent researchers on similar tasks. Using state of the art methods, theories and algorithms of behavioral economics, machine learning and deep learning we have advanced research in this area and contributed to increased understanding of human behavior prediction.