Lightly Supervised Machine Learning for Wireless Signals

Josh Sanz

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2024-227
December 20, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-227.pdf

Modern wireless communication systems face unprecedented challenges in managing finite spectrum resources while meeting growing demands for data and connectivity. This dissertation explores how machine learning techniques with reduced supervision requirements can address these challenges through three complementary approaches. First, I demonstrate that two radio agents with minimal shared assumptions can learn compatible modulation schemes through cooperative interaction, enabling communication without explicit protocol design. Careful experimentation, including simulation and implementation on software-defined radios, shows that while reduced supervision increases learning time, agents can still achieve near-optimal performance. Second, I develop techniques for automatic calibration and metadata generation in distributed spectrum sensing networks using signals of opportunity as a form of environmental supervision. These techniques enable verification of sensor characteristics like field of view and location without manual intervention, facilitating trustworthy large-scale deployments. Finally, I propose using generative models to allow the sharing of wireless datasets while preserving privacy, addressing a key barrier to advancing wireless machine learning research. The unifying theme is the development of techniques that minimize required human supervision while maintaining robust performance, enabling more autonomous and scalable wireless systems. This research is a step toward cognitive radio networks that can adaptively and cooperatively manage spectrum resources with reduced human oversight.

Advisor: Anant Sahai

"; ?>

BibTeX citation:

@phdthesis{Sanz:EECS-2024-227,
    Author = {Sanz, Josh},
    Title = {Lightly Supervised Machine Learning for Wireless Signals},
    School = {EECS Department, University of California, Berkeley},
    Year = {2024},
    Month = {Dec},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-227.html},
    Number = {UCB/EECS-2024-227},
    Abstract = {Modern wireless communication systems face unprecedented challenges in managing finite spectrum resources while meeting growing demands for data and connectivity. This dissertation explores how machine learning techniques with reduced supervision requirements can address these challenges through three complementary approaches. First, I demonstrate that two radio agents with minimal shared assumptions can learn compatible modulation schemes through cooperative interaction, enabling communication without explicit protocol design. Careful experimentation, including simulation and implementation on software-defined radios, shows that while reduced supervision increases learning time, agents can still achieve near-optimal performance. Second, I develop techniques for automatic calibration and metadata generation in distributed spectrum sensing networks using signals of opportunity as a form of environmental supervision. These techniques enable verification of sensor characteristics like field of view and location without manual intervention, facilitating trustworthy large-scale deployments. Finally, I propose using generative models to allow the sharing of wireless datasets while preserving privacy, addressing a key barrier to advancing wireless machine learning research. The unifying theme is the development of techniques that minimize required human supervision while maintaining robust performance, enabling more autonomous and scalable wireless systems. This research is a step toward cognitive radio networks that can adaptively and cooperatively manage spectrum resources with reduced human oversight.}
}

EndNote citation:

%0 Thesis
%A Sanz, Josh
%T Lightly Supervised Machine Learning for Wireless Signals
%I EECS Department, University of California, Berkeley
%D 2024
%8 December 20
%@ UCB/EECS-2024-227
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-227.html
%F Sanz:EECS-2024-227