Information Procuration via Adaptive Algorithms

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

The Principal Investigator (PI) formulates the information procuration problem of converting unstructured data into structured information as one of using limited resources (such as processing time and collection costs) among several available strategies for information acquisition, extraction, collation and aggregation in a sequential and adaptive manner. The proposal aims to build a Markov decision process (MDP) for which both the states and the rewards will be learned, and from which an optimal adaptive strategy for effective information procuration will be extracted. Recent methods for designing adaptive strategies for multi-armed bandit problems and budgeted learning approaches by the PI will be extended for this purpose, as well as techniques from inverse reinforcement learning. Moreover, given the intended size of the application data sets, the focus will be on on scalable algorithms for these problems. Due to the centrality of the problem, new approaches to making better sense of unstructured data will have much impact both in terms of developing new methods and in practice. The proposed synthesis of methods from Operations Research, Approximation Algorithms and Machine Learning is novel in this context. This proposal will increase the cross-fertilization of ideas between Operations Research and Machine Learning, via a collaboration team formed at this intersection.