Towards Reduced Conservatism of Model Predictive Control for Cyber-Physical Interconnected Systems : Contributions in Multi-Tube Formulations, Output-Feedback, and Hierarchical-Distributed Control

Model predictive control (MPC) is by now a widely used and well-established control method. It has become a door opener for many fields of applications, ranging from robotics, autonomous systems, and drones to power network control, aerospace, and process engineering. Its appeal lies in the systematic ability to handle constraints while exploiting predictive models for performance. At the same time, MPC offers a framework that combines theoretical rigor with flexibility. This makes it not only a cornerstone of control theory and control practice but also a natural foundation for modern learning-based approaches, where robust MPC formulations can provide the guarantees that are otherwise missing when using datadriven or machine learning models in control. Despite these strengths, uncertainty remains a central challenge — models approximate reality imperfectly, and systems face external disturbances, parameter variations, and uncertain measurements. Robust tube-based MPC addresses these challenges by ensuring constraint satisfaction and stability through set-based formulations. However, traditional approaches suffer from significant conservatism: the requirement to account for all possible uncertainty realizations leads to restrictive control actions, reduced performance, and sometimes infeasibility. This dissertation develops novel methods that systematically reduce this conservatism while maintaining rigorous theoretical guarantees. The thesis presents four main contributions. First, it introduces multi-tube robust MPC, which employs multiple tubes with different control gains instead of a single tube. By optimally interpolating between these tubes online, the approach achieves larger feasible regions and improved performance with only a marginal increase in computational complexity — the resulting optimization problems remain convex quadratic programs similar to standard MPC. Second, it develops advanced output-feedback robust MPC that reduces conservatism through unified error bounding. Rather than separately bounding estimation and prediction errors, a combined error dynamics approach using a single set provides tighter bounds. This innovation significantly enlarges the domain of attraction while maintaining recursive feasibility and stability guarantees. Third, the thesis addresses distributed control of interconnected systems through cyclic horizon formulations. This enables structured terminal costs and constraints that would be impossible with fixed horizons, facilitating coordination between subsystems. Fourth, using a cyclic horizon a framework is demonstrated through a comprehensive hierarchical planning and control system for autonomous vehicles, where different operating modes (e.g., slow-precise vs. fast-uncertain) are selected online to reduce conservatism while guaranteeing collision avoidance. All methods are rigorously analyzed, proving recursive feasibility, constraint satisfaction, and stability (including robust stability for uncertain systems). The theoretical developments are validated through extensive simulations spanning academic examples to realistic applications including quadcopters, chemical reactors, and mechanical systems. Importantly, the thesis provides a balanced assessment, acknowledging both capabilities and limitations. The contributions form a coherent framework advancing set-based robust MPC beyond traditional boundaries. The developed methods maintain the computational tractability essential for real-time implementation — most resulting in standard convex optimization problems solvable with existing tools. This work provides foundations for scalable control of networked cyber-physical systems, integration with learning-based approaches, and deployment in embedded systems where computational resources are limited. By systematically addressing conservatism across multiple dimensions — tube formulations, output feedback, and system architectures — this dissertation significantly extends the practical applicability of robust MPC while preserving the theoretical rigor essential for safety-critical applications.