Differences

This shows you the differences between two versions of the page.

--- courses:hcai:lab_uncertainty [2026/11/01 11:48] – created kkt
+++ courses:hcai:lab_uncertainty [2026/11/01 19:41] (current) – [Learn more!] kkt
@@ Line 15: / Line 15: @@
 ==== Q&A ====
-  - FIXME
+  - Can more evidence make a model less reliable?
+  - Why can confirming one cause make another cause less likely?
+  - Why does variable elimination order matter if the posterior remains the same?
+  - Can a correctly fitted Bayesian network still give misleading probabilities?
 ==== Notebook ====
@@ Line 25: / Line 28: @@
 ==== Challenge ====
-FIXME
+Work in groups of **2-4** using the standard **[[https://www.bnlearn.com/documentation/man/alarm.html|ALARM Bayesian network]]** and the official **[[https://pgmpy.org/|pgmpy]]** parameter-learning workflow.
+The network represents dependencies between variables used in patient monitoring. Instead of building a Bayesian network from scratch, you will load an existing benchmark model, investigate its behavior, generate observations, and determine how limited or biased data affect probabilistic inference.
+Begin with the official pgmpy example: [[https://pgmpy.org/examples/Parameter_Learning_Discrete_BN.html|Parameter Learning in Discrete Bayesian Networks]] \\
+Use the benchmark model as in the example: <code python>
+from pgmpy.example_models import load_model
+alarm_model = load_model("bnlearn/alarm")
+</code>
+__**Your investigation:**__ \\
+Choose one meaningful query variable and a small set of related evidence variables. Then:
+  * inspect the relevant part of the network;
+  * define at least four evidence scenarios;
+  * predict the direction of each posterior change before running inference;
+  * compare at least two variable-elimination orders;
+  * generate datasets of different sizes from the reference model;
+  * learn CPDs using maximum-likelihood and Bayesian estimation;
+  * compare the learned posteriors with the reference model;
+  * introduce one controlled data problem, such as selection bias or an underrepresented parent configuration;
+  * explain one limitation of your conclusions.
+Start the challenge during the lab and continue it as part of the **self-study regarding classes**. Selected groups may share their findings at the beginning of the next lab, and particularly insightful analyses may receive bonus points.
 ==== Learn more! ====
-  * FIXME
+  * [[https://seeing-theory.brown.edu/bayesian-inference/index.html|Seeing Theory: Bayesian Inference]] -- interactive visualizations of Bayes’ theorem, likelihood, and prior-to-posterior updating
+  * [[https://pgmpy.org/examples.html|pgmpy Example Notebooks]] -- examples extending the laboratory toward structure learning, approximate inference, simulation, dynamic Bayesian networks, junction trees, and causal inference.
+  * [[https://www.bnlearn.com/bnrepository/|Bayesian Network Repository]] -- standard benchmark networks of different sizes and domains, including ASIA, ALARM, CHILD, INSURANCE, HAILFINDER, and PATHFINDER.
+  * [[https://online.bayesserver.com/|Bayes Server Online]] -- an interactive environment for exploring Bayesian networks and observing how evidence changes probability distributions.