AN ADVERSARIAL BANDIT FRAMEWORK FOR PERIMETER PROTECTION

This research develops a principled method to efficiently collect intelligence from strategic targets in the absence of prior information. The resulting algorithms can be employed to geo-locate and characterize targets who attempt to avoid detection while crossing a perimeter. Our model considers the following scenario: agents operate in separate sectors of a perimeter, and a searcher attempts to interdict their activities. In every time period, each agent receives a demand of activities to perform; this demand is random, with a distribution that is known to the agent but not to the searcher. Each agent decides how many of the requested activities to perform, knowing that if the searcher looks into the agent’s sector a penalty proportional to the number of activities is incurred. From the searcher’s perspective, the objective is to detect as many malevolent activities as possible. To decide which sector to search, the searcher applies an adversarial bandit algorithm (like EXP3), which samples from a distribution that assigns more weight to the most lucrative sector so far. Knowing this, the agents balance their goal of performing activities versus the risk of getting caught. This narrative applies to perimeter protection settings more generally. The ultimate goal of the research is to design algorithms that optimally balance the exploration and exploitation of the data in the presence of strategic agents.http://archive.org/details/anadversarialban1094564842Outstanding ThesisLieutenant Commander, Republic of Korea NavyApproved for public release; distribution is unlimited