For any multi-agent system to function, its agents must first be able to reliably agree on critical information. This fundamental problem of reaching an agreement is known as consensus. However, in real-world deployments, communication can be unreliable, and some agents may be faulty or even actively malicious (adversarial).

My research in this area focuses on creating trustworthy and resilient consensus protocols. I design and analyze algorithms that guarantee the honest agents in a network can reach a correct consensus, even in the presence of intermittent failures and malicious transmissions from adversarial agents. By building this layer of trust and resilience at the foundational level of communication, we can ensure the safety, reliability, and overall success of complex collaborative missions.

I bridge the gap between classic game theory and modern machine learning through Multi-Agent Reinforcement Learning (MARL). While traditional game theory often assumes agents have perfect information, MARL allows agents to learn optimal strategies through trial and error, adapting to the environment and the behavior of other agents over time.

A significant part of my research involves designing novel policy gradient algorithms for networked agents in Markov potential games. This approach empowers agents to learn complex, state-dependent policies directly from experience, without needing a central coordinator. By integrating policy gradient methods with the structural guarantees of potential games, we can develop learning dynamics that are both scalable and provably stable for large-scale multi-agent systems.

At its core, my research uses the game theory to model strategic interactions between autonomous agents. I focus on scenarios where agents are self-interested yet must coordinate to achieve a collective goal. A key area of my work is on potential games, a special class of games where the incentives of individual agents can be aligned with a global objective function. This property allows us to design decentralized algorithms that are guaranteed to converge to a mutually beneficial and stable outcome, known as a Nash Equilibrium.

My work develops learning algorithms like decentralized fictitious play and best-response dynamics, where each agent learns by observing the past actions of its neighbors. A critical aspect of my research is accounting for real-world constraints, such as agents operating over time-varying communication networks. This ensures that the solutions are robust and applicable to dynamic settings, such as mobile robotic teams performing complex tasks like target assignment.