A Machine Learning Approach to Modeling Dynamic Decision-Making in Strategic Interactions and Prediction Markets

摘要

My dissertation lies at the intersection of computer science and the decision sciences. With psychology and sociology, I'm interested in models where social and cognitive factors influence human decisions, especially in social dilemmas and out-of-equilibrium dynamics such as learning and adaptation. With microeconomics and game theory, I realize that modeling human behavior as an attempt to maximize individual well-being is useful. I combine theoretical insights from the decision sciences with computational methods to understand and predict human behavior.;My work is distinct from most decision science research in two respects: (1) I emphasize prediction, and (2) I am not attempting to make simple economic decision models better describe behavior. First, the majority of economic and behavioral science research focuses on either describing in-sample phenomena or testing theories that posit causal relationships among theoretical constructs [1, 2]. The primary difference between prediction and causality research approaches derives from the unit of interest -- causal explanation is directly concerned with theoretical population-level constructs, while prediction is directly concerned with sampled data. My research demonstrates that data-driven models with a prediction focus can be strategically designed and implemented to inform theory. As for the second point of divergence, I am not part of what constitutes most of the field of behavioral economics: the "subjective expected utility repair program" [6]. This is the active line of research adding psychological parameters to the subjective expected utility model [7] to allow it to better fit behavioral data [8].;My overarching modeling goal for my dissertation is to maximize generalization -- some function of data and knowledge -- from one sample, with its observations drawn independently from the distribution D, to another sample drawn independently from D,2 while also obtaining interpretable insights from the models. The processes of collecting relevant data and generating features from the raw data impart substantive knowledge into predictive models (and the model representation and optimization algorithms applied to those features contain methodological knowledge). I combine this knowledge with the data to train predictive models to deliver generalizability, and then investigate the implications of those models with simulations systematically exploring parameter spaces. The exploration of parameter space provides insights about the relationships between key variables.