Control and Learning for Autonomous Systems

<p dir="ltr">Rooted in the idea that complex objectives are best achieved through cooperation, multi-agent systems (MAS) now underpin applications across engineering, optimization, economics, and the social sciences. The central challenge is to realize global goals using only local coordination and constrained communication. This makes collaboration and communication strategies a first-class concern for developing algorithms for MAS. Alongside these coordination issues, two major application pillars shape the field: control, which covers autonomous vehicles, robotics, and process systems, and learning, which spans distributed machine learning, cooperative inference, and swarm intelligence.</p><p dir="ltr">This thesis examines three intertwined facets: collaboration and communication, control, and learning, and contributes new frameworks and algorithms in each. On collaboration and communication, our work introduces edge agreement, a framework that generalizes consensus from global agreement to heterogeneous pairwise constraints on graph edges, with classical consensus recovered as a special case. We also study and analyze projection-based consensus where agents see only low-dimensional views of neighbor states, and we characterize convergence through an extended representation that couples sensing and topology. For clustered MAS, we design a leaderless distributed method to solve large linear systems when each agent holds only a local block, avoiding cluster aggregators and improving scalability. On control, in the thesis we develop neighboring extremal optimal control (NEOC) for closed-loop policies, deriving sensitivity relations that quantify how optimal feedback changes under small perturbations to dynamics or costs. We propose performance index shaping, which links cost parameters to the resulting feedback law to provide interpretable tuning with stability insight. We also present a unified safe region formation control method for second-order agents that enforces region reaching, collision avoidance, connectivity preservation, actuator bounds, and limited velocity information. On learning, in the thesis, we design a fully distributed strategy for kernelized bandit optimization with heterogeneous rewards, balancing exploration and exploitation through local uncertainty and neighbor exchanges. We also study federated learning through a control lens and propose FedProject, a projection-regularized proximal update that mitigates client drift under non-IID data while remaining communication-efficient.</p>