Learning to Play Stackelberg Security Games

摘要: As discussed in previous chapters, algorithmic research on Stackelberg Security Games has had a striking real-world impact. But an algorithm that computes optimal strategy for the defender can only be as good game it receives input, and if is inaccurate model of reality then output will likewise flawed. Consequently, researchers have introduced Bayesian frameworks capture uncertainty using probability distribution over possible games. Others assumed unknown parameters lie within known intervals. These approaches are Chapter 17 this book [17]. In chapter, we present alternative, learning-theoretic approach dealing with security order to paint cohesive picture, focus one type uncertainty: attacker utilities. Learning take place repeated game, where gathers information about purely by observing attacker’s responses mixed strategies played defender. more detail, wish learn without any initial utility function (Section 1); when given types 2); faced sequence attackers 3). each section present, some generality, relevant techniques: optimization membership queries, Monte Carlo tree search, no-regret learning, respectively. Section 4 briefly discuss additional work at intersection machine learning