Chinmay Maheshwari

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-138

June 18, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-138.pdf

Autonomous AI technologies are increasingly embedded in critical societal systems, including robotics, transportation, logistics, and energy—where they enable large-scale, data-driven decision-making. While recent advances have enabled autonomous agents to perform effectively in isolated or structured environments, a fundamental open challenge is to integrate such agents into dynamic, uncertain, and resource-constrained multi-agent environments where they must learn and interact strategically with other autonomous systems and with humans. The emerging outcomes not only impact the individual utility but also impacts societal efficiency, equity and safety.

This dissertation addresses the design and analysis of intelligent autonomous agents in such multi-agent societal settings. It is motivated by two central questions: (1) How can we design learning and decision-making algorithms that allow autonomous agents to act rationally and strategically in the presence of other agents? (2) How can we ensure that the collective outcomes of such agent interactions align with broader societal goals such as efficiency, equity, and safety?

To answer these questions, the dissertation introduces new theoretical, algorithmic, and computational frameworks for multi-agent learning, decision-making, and design of multi-agent interactions in societal systems. These contributions are organized into four parts, each grounded in application domains that highlight key challenges and propose novel solutions.

Part I focuses on learning in general-sum Markov games, which model multi-agent interactions in uncertain, dynamic environments. Unlike classical control or RL settings that assume either fully cooperative or fully adversarial interactions, many real-world systems exhibit a mix of cooperative and competitive behavior. To address this, we propose a new theoretical framework of \emph{Markov near-potential games}, which approximates the underlying multi-agent interaction using a potential game. We leverage this framework to design and analyze multi-agent learning algorithms. Specifically, we use it to design real-time, high-performance strategies for autonomous multi-car racing that outperform several existing baselines. Additionally, we use the framework to characterize the long-run outcomes of interactions between decentralized reinforcement learning algorithms, with a focus on actor-critic methods.

Part II examines strategic learning under competition induced due to shared resource and infrastructure constraints, including settings with congestion. The focus is on domains such as transportation networks and two-sided matching markets, where agents compete over scarce, congestible resources. This part introduces learning dynamics that achieve desirable performance guarantees—such as low regret and equilibrium convergence—even when agents adapt based on local observations and uncertain feedback.

Part III shifts from agent-level optimization to designing mechanisms to align strategic agent behavior with societal objectives. A key challenge here is that agents may respond strategically to deployed mechanisms, leading to distribution shifts, while designers often lack access to private agent preferences. This part proposes data-driven methods for design of societal mechanisms that remain robust to strategic behavior and result in socially beneficial outcome. We highlight applications in design of congestion pricing on road networks and design of data-driven online services.

Part IV explores market design for the emerging AAM—a future mobility paradigm involving UAVs and air taxis operating in low-altitude urban airspace. Given the decentralized and adaptive nature of AAM systems, traditional centralized air traffic control methods are inadequate. This part introduces market-based mechanisms for allocating trajectories to UAVs with potentially heterogeneous preferences that ensure safety, fairness, and efficiency.

Advisors: S. Shankar Sastry

---

BibTeX citation:

@phdthesis{Maheshwari:EECS-2025-138,
    Author= {Maheshwari, Chinmay},
    Title= {Dynamic Multi-agent Autonomous Systems for Societal Transformation},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Jun},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-138.html},
    Number= {UCB/EECS-2025-138},
    Abstract= {Autonomous AI technologies are increasingly embedded in critical societal systems, including robotics, transportation, logistics, and energy—where they enable large-scale, data-driven decision-making. While recent advances have enabled autonomous agents to perform effectively in isolated or structured environments, a fundamental open challenge is to integrate such agents into dynamic, uncertain, and resource-constrained multi-agent environments where they must learn and interact strategically with other autonomous systems and with humans. The emerging outcomes not only impact the individual utility but also impacts societal efficiency, equity and safety.

This dissertation addresses the design and analysis of intelligent autonomous agents in such multi-agent societal settings. It is motivated by two central questions:
(1) How can we design learning and decision-making algorithms that allow autonomous agents to act rationally and strategically in the presence of other agents?
(2) How can we ensure that the collective outcomes of such agent interactions align with broader societal goals such as efficiency, equity, and safety?

To answer these questions, the dissertation introduces new theoretical, algorithmic, and computational frameworks for multi-agent learning, decision-making, and design of multi-agent interactions in societal systems. These contributions are organized into four parts, each grounded in application domains that highlight key challenges and propose novel solutions. 

Part I focuses on learning in general-sum Markov games, which model multi-agent interactions in uncertain, dynamic environments. Unlike classical control or RL settings that assume either fully cooperative or fully adversarial interactions, many real-world systems exhibit a mix of cooperative and competitive behavior.
To address this, we propose a new theoretical framework of \emph{Markov near-potential games}, which approximates the underlying multi-agent interaction using a potential game. We leverage this framework to design and analyze multi-agent learning algorithms. Specifically, we use it to design real-time, high-performance strategies for autonomous multi-car racing that outperform several existing baselines. Additionally, we use the framework to characterize the long-run outcomes of interactions between decentralized reinforcement learning algorithms, with a focus on actor-critic methods.

Part II examines strategic learning under competition induced due to shared resource and infrastructure constraints, including settings with congestion. The focus is on domains such as transportation networks and two-sided matching markets, where agents compete over scarce, congestible resources. This part introduces learning dynamics that achieve desirable performance guarantees—such as low regret and equilibrium convergence—even when agents adapt based on local observations and uncertain feedback.

Part III shifts from agent-level optimization to designing mechanisms to align strategic agent behavior with societal objectives. A key challenge here is that agents may respond strategically to deployed mechanisms, leading to distribution shifts, while designers often lack access to private agent preferences. This part proposes data-driven methods for design of societal mechanisms that remain robust to strategic behavior and result in socially beneficial outcome. We highlight applications in design of congestion pricing on road networks and design of data-driven online services.

Part IV explores market design for the emerging AAM—a future mobility paradigm involving UAVs and air taxis operating in low-altitude urban airspace. Given the decentralized and adaptive nature of AAM systems, traditional centralized air traffic control methods are inadequate. This part introduces market-based mechanisms for allocating trajectories to UAVs with potentially heterogeneous preferences that ensure safety, fairness, and efficiency.},
}

EndNote citation:

%0 Thesis
%A Maheshwari, Chinmay 
%T Dynamic Multi-agent Autonomous Systems for Societal Transformation
%I EECS Department, University of California, Berkeley
%D 2025
%8 June 18
%@ UCB/EECS-2025-138
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-138.html
%F Maheshwari:EECS-2025-138