Full report: report_berlinger_xu.pdf
Given a training set of 250 3x3 bi-matrix games, we attempted to infer the behavior of the row player and make predictions on a test set accordingly. Evaluation metrics include the quadratic distance of the frequency distribution (Q) and the accuracy in predicting the most frequently chosen action (A). We attempted behavioral inference and prediction using (i) traditional models such as Nash Equilibrium or Level-k, (ii) Machine Learning strategies including Linear Regression and Gradient Boost, as well as (iii) Hybrid models that combine the former two approaches.