Ingebedde leer- en optimalisatiealgoritmen voor interactiebewuste modelgebaseerde predictieve controle

As robots increasingly operate in public spaces shared with humans and other robots, accounting for the reactive behavior of surrounding agents becomes crucial for safe and effective autonomous operations. This new type of problem brings various challenges, including the need for data-driven modeling of interaction dynamics, and efficient control algorithms under this new setup, etc. Among these challenges, a crucial theoretical bottleneck is the nonconvex stochastic optimization problem with decision-dependent uncertainty arising from the interactive nature of the problem, which is a long-standing issue in the field of stochastic optimization. This thesis aims to tackle these challenges arising from the interactive settings with a specific focus on this theoretical bottleneck. To this end, we decompose the problem into three smaller tasks that can be addressed systematically while accounting for their interconnections: (i) developing expressive models that are suitable for downstream controller design; (ii) creating control formulations tailored for decision-dependent uncertainty in the interactive setting; (iii) designing efficient algorithms exploiting the mathematical structure of both modeling and control problems. Following this approach, the contribution of this thesis is in threefold: (i) We proposed a generalized mixture-of-expert model to capture the interactive behavior accompanied by a learning framework to learn parameters from data; (ii) We developed a risk-sensitive optimal control framework tailored for interaction-aware planning and control, which accounts for the strategic behavior of other agents in the interactive settings; (iii) We developed majorization-minimization-based algorithms for both the parameter learning problem and the risk-sensitive optimal control problem, which utilize the problem structure to make them more computationally efficient. Through autonomous navigation applications, we demonstrated the effectiveness of proposed approaches while establishing foundational techniques applicable to the broader class of interaction-aware control problems.