A multi-armed bandit approach to superquantile selection
Hepworth, Adam J.
Szechtman, Roberto S.
Atkinson, Michael P.
MetadataShow full item record
We study a resource allocation problem in an intelligence setting. The intelligence cycle is comprised of three phases: collection, processing, and analysis. Enhanced efficiency within the first two stages directly impacts the number and types of important items that are considered by analysts, increasing the frequency of the most important documents that are reviewed. The dilemma here is that an analyst needs to quickly determine which sources to investigate, in order to provide meaningful analysis to a request for information with, a concrete deadline. Initially, the value of each source is unknown; so, too, is the probabilistic nature of the value derived from each item. Generally, more sources and documents are available to be considered within a limited time frame than could be ever analyzed, compounding the complexity of this problem. Our goal is to efficiently find the source that produces the largest fraction of relevant items with respect to a request for information. By "efficiently," we mean that the analyst balances exploration versus exploitation of the different sources judiciously. As such, the theoretical framework for this problem is that of a multi-armed bandit, a classic iterative decision learning process. This thesis presents a new approach to identifying the optimal arm(s) of a multi-arm bandit with the largest or smallest quantile or superquantile risk, under a loss constraint. This problem is not only important in intelligence applications, but in marketing and finance. We extend the existing theoretical framework of dealing with quantiles to a novel situation with estimators of conditional expectations over an unknown quantile. Two sequential elimination algorithms are developed that select the most important source for a given constraint level, sampling from the arm(s) with the largest conditional expectation over a quantile.
Approved for public release; distribution is unlimited
Showing items related by title, author, creator and subject.
Political and budgetary oversight of the Ukranian intelligence community processes, problems and prospects for reform Petrov, Oleksii (Monterey, California. Naval Postgraduate School, 2007-09);This thesis addresses the problem of providing policy and budget oversight of Ukrainian intelligence organizations in accordance with norms and practices developed in contemporary Western democracies. As Ukraine continues ...
Naval Postgraduate School Center for Homeland Defense and Security (CHDS) (Monterey, California. Naval Postgraduate SchoolCenter for Homeland Defense and Security, 2007-09);September 2007. Six years after the attacks of 9/11, the practice and discipline of homeland defense and security have evolved and matured, moving into an era of self-evaluation. The essays and articles in Volume III, Issue ...
Costica, Yinon (Monterey, California. Naval Postgraduate School, 2010-12);The intelligence making process, often described as the intelligence cycle, consists of phases. Congestion may be experienced in phases that require time consuming tasks such as translation, processing and analysis. To ...