Learning Adversary Behavior in Security Games: A PAC Model Perspective. (arXiv:1511.00043v1 [cs.AI])

Recent applications of Stackelberg Security Games (SSG), from green crime to urban crime, have employed machine learning tools to learn and predict adversary behavior using available data about defender-adversary interaction.We commit to an approach of directly learning the response function of the adversary and initiate a formal study of the learnability guarantees of the approach. We make three main contributions: (1) we formulate our approach theoretically in the PAC learning framework, (2) we analyze the PAC learnability of the known parametric SUQR model of bounded rationality and (3) we propose the first non-parametric class of response functions for SSGs and analyze its PAC learnability. Finally, we conduct experiments and report the real world performance of the learning methods.

from cs.AI updates on arXiv.org http://ift.tt/1MDqfrg
via IFTTT