Causal Analysis and Decision Support

Seven biweekly blocks build the theory with worked derivations, each paired with a project that serves as its laboratory component. The course is organised in two arcs: advanced causal analysis (Blocks 1–4) and the engineering of decision-support systems (Blocks 5–7). 

Block 1 — Identification: how far does the do-calculus reach? (Weeks 1–2) 

A rapid recap of structural causal models, the causal ladder, and the three rules of the do-calculus, then the complete theory of identification, the completeness of the do-calculus, and the identification of conditional and counterfactual queries. Adjustment beyond the back door — front-door and general adjustment — and the choice among valid adjustment sets by statistical efficiency. 

Block 2 — Causal discovery I: structure from observation (Weeks 3–4) 

Recovering causal structure when it is not given. Constraint-based discovery and score-based discovery; Markov-equivalence classes and their representation; latent confounders and selection bias through maximal ancestral graphs ; and the assumptions on which all of this rests — the causal Markov condition, faithfulness, and causal sufficiency. 

Block 3 — Causal discovery II: asymmetries, interventions, and priors (Weeks 5–6) 

Going beyond the equivalence class. Functional causal models that exploit distributional asymmetry — linear non-Gaussian, additive-noise, and post-nonlinear models; Bayesian structure learning with a posterior over graphs and model-averaged effects; active experimental design that selects interventions to orient edges; and the use of the literature, ontologies, and large language models as informative structure priors. 

Block 4 — Effect estimation, sensitivity, and partial identification (Weeks 7–8) 

Turning an identified estimand into a defensible number. Inverse-probability weighting, outcome regression, and doubly-robust and targeted estimators; instrumental variables; double / debiased machine learning for high-dimensional nuisance models; sensitivity analysis to unmeasured confounding; and partial identification — informative bounds when ignorability does not hold. 

Block 5 — Decision support I: influence diagrams, value of information, and DSS architecture (Weeks 9–10) 

From a single decision to a system. Multi-stage influence diagrams and limited-memory influence diagrams and their solution for optimal policies; the value of information and the value of control at scale; and the anatomy of a decision-support system — knowledge base, inference engine, elicitation, explanation, and interface — together with the knowledge engineering and calibration of elicited probabilities and utilities. 

Block 6 — Decision support II: sequential decisions, explanation, and recourse (Weeks 11–12) 

Acting under partial observability, and accounting for the action afterwards. From Markov decision processes to partially observable Markov decision processes; belief-state planning, point-based value iteration, and online planning, connecting back to Monte-Carlo Tree Search; and the explanatory side of decision support — most-probable-explanation and MAP inference, counterfactual and contrastive explanations, and actionable recourse, with calibrated deference to the human. 

Block 7 — Decision support III: LLM-augmented, multi-agentBuilding directly on Foundations of Artificial Intelligence, this optional course advances the unified probabilistic, causal, and decision-theoretic framework toward two applied frontiers: rigorous causal analysis and the engineering of decision-support systems. On the causal side it moves from the do-calculus to complete identification, from a single Bayesian network to causal discovery and its equivalence classes, and from textbook adjustment to Objectives. Building Foundations of Intelligence, course students carry out causal to design, explain, and support It identification, robust estimation the side, and partially sequential decision-explanation, and deployment the decision keeping method with provable limit.      outcomes. the end, can:      whether a query is derive estimand, certify non-run and read methods, Markov-equivalence CPDAGs, the their estimate counterfactual with targeted, and methods, and sensitivity