Learning and Decision Making Under Observer Effects

This lecture by Sarah Dean from Cornell University explores the challenges of learning and decision making in systems with "observer effects," where measurements and underlying quantities become interdependent. Discover how decisions impact both system states and what information can be observed—a critical consideration for developing reliable algorithms across robotics and personalized recommendation systems. The talk focuses on partially observed dynamical systems with linear state transitions and bilinear observations, drawing inspiration from research on learning and control for linear systems. Learn about identification procedures involving heavy-tailed and dependent covariates, including finite data error bounds and sample complexity analysis for randomly designed inputs. Examine why optimal control problems with quadratic cost objectives differ from standard linear quadratic Gaussian control, explaining why separation principle-based controllers can maximize costs instead of minimizing them under certain conditions. Based on joint work with Yahya Sattar, Sunmook Choi, Yassir Jedra, and Maryam Fazel, this presentation from the Theoretical Aspects of Trustworthy AI program offers valuable insights into this complex interdisciplinary field.