APM_53452_EP - Constrained (reinforcement) learning (2026-2027)

Requirements are of growing interest in ML as evidenced by the advancement towards designing fair, robust, and/or safe AI systems. This course introduces the theoretical and algorithmic foundations of constrained learning, an emerging technique to organically incorporate requirements when training of AI models. It will cover constrained optimization and duality prerequisites, showcasing the main challenges faced by constrained learning. It then develops the main generalization guarantees and algorithms for supervised learning, showcasing their use in fairness, robustness, invariance, and scientific applications. Finally, it derives a parallel theory and algorithms for reinforcement learning problems, using it to show the limitations of penalty-based techniques in this sequential decision making setting.