Doctoral Dissertations

Date of Award


Degree Type


Degree Name

Doctor of Philosophy


Electrical Engineering

Major Professor

Husheng Li

Committee Members

Hairong Qi, Xueping Li, Husheng Li, Arun Padakandla


The intersection of communication and machine learning is attracting increasing interest from both communities. On the one hand, the development of modern communication system brings large amount of data and high performance requirement, which challenges the classic analytical-derivation based study philosophy and encourages the researchers to explore the data driven method, such as machine learning, to solve the problems with high complexity and large scale. On the other hand, the usage of distributed machine learning introduces the communication cost as one of the basic considerations for the design of machine learning algorithm and system.In this thesis, we first explore the application of machine learning on one of the classic problems in wireless network, resource allocation, for heterogeneous millimeter wave networks when the environment is with high dynamics. We address the practical concerns by providing the efficient online and distributed framework. In the second part, some sampling based communication-efficient distributed learning algorithm is proposed. We utilize the trade-off between the local computation and the total communication cost and propose the algorithm with good theoretical bound. In more detail, this thesis makes the following contributionsWe introduced an reinforcement learning framework to solve the resource allocation problems in heterogeneous millimeter wave network. The large state/action space is decomposed according to the topology of the network and solved by an efficient distribtued message passing algorithm. We further speed up the inference process by an online updating process.We proposed the distributed coreset based boosting framework. An efficient coreset construction algorithm is proposed based on the prior knowledge provided by clustering. Then the coreset is integrated with boosting with improved convergence rate. We extend the proposed boosting framework to the distributed setting, where the communication cost is reduced by the good approximation of coreset.We propose an selective sampling framework to construct a subset of sample that could effectively represent the model space. Based on the prior distribution of the model space or the large amount of samples from model space, we derive a computational efficient method to construct such subset by minimizing the error of classifying a classifier.


Portions of this document were previously published in journal.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."