H.B. Keller Colloquium

Control and machine learning are two high-impact areas. The developments of next generation intelligent systems such as self-driving vehicles, humanoid robotics, smart buildings, and automated healthcare require a rapprochement of these two areas. This talk will cover the fundamental connections between learning and control, and also discuss how generative AI is changing our perspectives on how these two fields are bridged. The first half of this talk focuses on some connections of learning and control that were built before the era of generative AI. We will discuss how to tailor control-theoretic tools such as quadratic constraints to unify the developments of learning algorithms and models. In addition, we will also borrow tools from modern optimization/learning theory to develop global convergence guarantees of policy-based reinforcement learning methods on important robust control benchmark problems such as H-infinity state-feedback synthesis. In the second half of the talk, we will focus on how generative AI shifts the research on interplay between control and machine learning. We will discuss our recent work on exploring the capabilities of state-of-the-art large language models (LLMs), such as GPT-4, Claude 3 Opus, and Gemini 1.0 Ultra, in solving control design problems. In particular, we introduce ControlBench, a benchmark dataset tailored to reflect the breadth, depth, and complexity of classical control design. Our analysis reveals the strengths and limitations of each LLM in the context of classical control, and our results imply that Claude 3 Opus has become the state-of-the-art LLM for solving undergraduate control problems. Our study serves as an initial step towards the broader goal of employing artificial general intelligence in control engineering. If time permits, we will also discuss two more papers on rethinking the concept of controllability in the context of generative AI.