Active and Dynamical Information Acquisition with Applications in Communication Systems
- Author(s): Ronquillo, Nancy
- Advisor(s): Javidi, Tara
- et al.
This dissertation investigates the task of quickly and accurately learning possibly time-varying information about a system by making binary and noisy measurements that yield insight into its current state. This problem arises in beam alignment for millimeter-wave (mmWave) communication where the active learning of the location of a transmitter relative to the receiver is necessary to establish communication. In the problem of spectrum sensing for cognitive radio, active learning of the spectrum occupancy is used for opportunistic communication. These applications motivate our study of the problem of searching for a target(s) among a discrete set of locations by probing different locations with the caveat that probing larger areas leads to more incurred noise. The problem of binary search has been extensively studied and we are inspired by information-theoretic principles to apply to practical problems in communication systems for enabling improved learning acquisition. Our methodology can be summarized by two central paradigms: active and sequential design of measurements, and dynamic tracking of a Bayesian posterior belief.
First, we cast the problem of searching for stationary, yet unknown, target locations as the problem of channel coding with state and feedback. We apply adaptive and sequential codes, i.e. measurement design, based on posterior matching to the problems of beam alignment and spectrum sensing to study these from a fundamental limit point of view. Our results characterize significant improvements in performance obtained by using adaptive measurements over non-adaptive ones, which is especially critical in the regimes of low signal-to-noise ratio. In the second half of this work, we generalize our work for learning time-varying measurement gains and dynamic target locations. We complement our strategies of active and sequential measurement design with simultaneous estimation of the measurement gains, which enables handling time-varying fading in mmWave beam alignment for example. Lastly, we enable handling stochastic mobility in mmWave beam alignment by incorporating predictive Bayesian filtering to dynamically evolve the posterior belief and by adaptively allocating pilots based on analysis of the mutual information and spectral efficiency. Combined, this dissertation moves towards solutions for the general and practical target search problem characterized by mobile and non-uniform targets.