Explore a comprehensive tutorial on recommendation systems framed as counterfactual policy learning. Delve into the conceptual frameworks behind state-of-the-art recommender systems, examining their underlying assumptions, methods, and limitations. Discover a new approach that views recommendation as a counterfactual policy learning problem. Learn about current approaches for building real-world recommender systems, including recommendation as optimal auto-completion of user behavior and as reward modeling. Examine theoretical guarantees addressing shortcomings of previous frameworks, and test associated algorithms against classical methods using RecoGym, an open-source recommendation simulation environment. Gain insights from industry experts on deep learning-based recommendation systems, causal inference in recommendation, and offline evaluation techniques.