Towards Algorithmic Advances for Solving Stackelberg Games: Addressing Model Uncertainties and Massive Game Scale-up

Abstract : This project opens up a brand new area of research that fuses two separate subareas of game theory: algorithmic game theory and behavioral game theory. More specifically, game-theoretic algorithms have been deployed by several security agencies, allowing them to generate optimal randomized schedules against adversaries who may exploit predictability. However, one key challenge in applying game theory to solving real-world security problems is the perfect rationality assumption of the players, which may not hold when dealing with human adversaries. Therefore, it is critical that we develop a new set of game-theoretic algorithms taking into account adversaries' bounded rationality. To that end, our accomplishments include: i)integrating mathematical models of human decision making based on Prospect Theory and Quantal Response into game-theoretic algorithms; ii)comprehensive experiments with human subjects which evaluates the effectiveness of these new algorithm showing improvement over the previous leading contender; iii) enhancing the efficiency of these gametheoretic algorithms, thus the use of these algorithms for computing security schedules in real-world settings.