Geng Zhao

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-161

August 15, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-161.pdf

Many modern applications of computing, ranging from school choices to epidemic model- ing, are fundamentally multi-agent in nature, involving interactions among strategic and often adaptive entities. This thesis presents theoretical methodologies for understanding macroscopic behaviors in such multi-agent systems and establishing insights that inform both inference and learning within such systems as well as the design and improvement of mechanisms and interventions.

Across a spectrum of settings, we showcase how mathematical analysis reveals novel under- standing on the structure and dynamics of various multi-agent systems shaped by strategic interactions and network structures. In Stackelberg games with information asymmetry, we demonstrate that expert follower behavior can paradoxically hinder the leader’s learning process, highlighting subtle challenges in multi-agent learning dynamics. In large two-sided matching markets, we introduce a notion of agent fitness under heterogeneous preferences and prove that stable outcomes exhibit universal statistical structure. Such structural characterization lends further insights on school choice systems, as we show that the widely used deferred acceptance mechanism is, with high probability, far from Pareto efficient. We then turn to the problem of estimating global parameters in networks based on local samples and propose a new robustness condition under which certain local estimations become viable. In another setting where network structure is crucial, we study the control of epidemics in networks with community structure, assessing the effectiveness of common intervention strategies, such as social distancing and travel restrictions. Finally, we highlight a simple token-based local policy for service systems such as kidney exchanges, and show that it effectively maintains stability of the system.

Together, the approaches we develop form a methodological toolkit for analyzing multi-agent systems with limited information, drawing on concepts from information theory, probability, random graph theory, and game theory. They demonstrate how mathematical analysis of local information and agent behavior can yield predictive and design-relevant insights into the global dynamics and performance of large-scale systems.

Advisors: Jiantao Jiao and Christian Borgs

---

BibTeX citation:

@phdthesis{Zhao:EECS-2025-161,
    Author= {Zhao, Geng},
    Title= {Perspectives on Multi-Agent Systems: From Characterization to Intervention},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-161.html},
    Number= {UCB/EECS-2025-161},
    Abstract= {Many modern applications of computing, ranging from school choices to epidemic model- ing, are fundamentally multi-agent in nature, involving interactions among strategic and often adaptive entities. This thesis presents theoretical methodologies for understanding macroscopic behaviors in such multi-agent systems and establishing insights that inform both inference and learning within such systems as well as the design and improvement of mechanisms and interventions.

Across a spectrum of settings, we showcase how mathematical analysis reveals novel under- standing on the structure and dynamics of various multi-agent systems shaped by strategic interactions and network structures. In Stackelberg games with information asymmetry, we demonstrate that expert follower behavior can paradoxically hinder the leader’s learning process, highlighting subtle challenges in multi-agent learning dynamics. In large two-sided matching markets, we introduce a notion of agent fitness under heterogeneous preferences and prove that stable outcomes exhibit universal statistical structure. Such structural characterization lends further insights on school choice systems, as we show that the widely used deferred acceptance mechanism is, with high probability, far from Pareto efficient. We then turn to the problem of estimating global parameters in networks based on local samples and propose a new robustness condition under which certain local estimations become viable. In another setting where network structure is crucial, we study the control of epidemics in networks with community structure, assessing the effectiveness of common intervention strategies, such as social distancing and travel restrictions. Finally, we highlight a simple token-based local policy for service systems such as kidney exchanges, and show that it effectively maintains stability of the system.

Together, the approaches we develop form a methodological toolkit for analyzing multi-agent systems with limited information, drawing on concepts from information theory, probability, random graph theory, and game theory. They demonstrate how mathematical analysis of local information and agent behavior can yield predictive and design-relevant insights into the global dynamics and performance of large-scale systems.},
}

EndNote citation:

%0 Thesis
%A Zhao, Geng 
%T Perspectives on Multi-Agent Systems: From Characterization to Intervention
%I EECS Department, University of California, Berkeley
%D 2025
%8 August 15
%@ UCB/EECS-2025-161
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-161.html
%F Zhao:EECS-2025-161