Shakespeare and Bayes are in a boat, fishing. Bayes is trying to figure out which net to cast when Shakespeare says: "loopy or not loopy? that is the question".

### Academic year 2018/2019:

November 22, 2018: The grades for the written exam as well as the final grades are available in BlackBoard's gradecenter. Blackboard also lists the scores for the different questions of the written exam. Final results for the course are also found in Osiris (see the Examination and grading page if you want to know whether or not you qualify for a retake).

To have a look at the marking of your exam, please contact one of the lecturers.

It seems like a lot of students were scared off or took way too much time for question 2, which affected both the scores on question 2 and 3. This was corrected for in computing the exam grade. Note that question 2b, where you had to analytically derive the form of the sensitivity function, basically amounts to applying marginalisation and the network factorisation: two of the very basic ingredients of the course, which everyone should be able to do. Similarly, question 3d, where you had to compute a conflict measure, in the given context basically just tested whether you were able to correctly apply probabilistic independence....again one of the basics of the course.

October 31, 2018: Here's a list of 4 remarks/announcements:

1. The grades for assignment D are available in Blackboard's Gradecenter. I will bring the marked assignments to class today (and Friday), so you can have a look at your own work.
2. Note that Friday is the last lecture, so take this opportunity to ask your remaining questions as well!
3. The mobile app for Blackboard is apparently showing lots of 0.0 for assignment grades; if this is the case for you, please go to Blackboard itself.
4. to prepare for the exam, I have annotated the course learning goals (also see page 2 and 3 of the studymanual):
• recognises and understands the strengths and weaknesses of probabilistic graphical models in general and Bayesian networks in particular;

This refers to global things concerning compactness of representation, complexity of inference, intuitiveness (or not?) of the graph and things like that. In the exam strenghts and weakness will not directly be asked for, but could for example be part of the motivation for a modelling choice presented to you.

• understands the relation between probabilistic independence and the graphical representations thereof, and is able to draw conclusions from this relation;

This means that you need to know how I-maps and D-maps encode independence relations, both in directed and undirected graphs.

This means more specifically that you need to know that a Bayesian network is an I-map and what that entails (that whatever d-separation statements we can read from the graph, we have equivalent independence statements in our independence relation for Pr, but not the other way around...we can therefore generally not conclude anything about dependence from a Bayesian network, unless the graph is also a D-map, hence a P-map).

You will not be asked to give definitions for active chains or d-separation or things like that, but you should be able to apply them even if this is not explicitly asked for.

• understands and is able to apply probabilistic inference in Bayesian networks;

This means you should be able to show how Pearl's algorithm works, with or without loop cutset conditioning. The loop cutset may be provided, or you have to establish it yourself (with or without heuristic).

• has knowledge and understanding of methods for constructing the Bayesian network graph for actual applications;

This concerns both approaches to handcrafting and learning the DAGs, as discussed in class.

• has knowledge and understanding of various methods for quantifying Bayesian networks, including parameterized models (noisy-or), together with their benefits and limitations;

This concerns the different sources of probabilistic information, how to reduce the number of probabilities requires, how to assess probabilities from data, or elicit them from domain experts and what the 'problems' with this can be

• understands and is able to apply techniques for evaluating the robustness and quality of a Bayesian network.

This concerns sensitivity analysis, the ways of making this computational feasible (sensitivity set and functional forms), the derivation and interpretation of sensitivity functions, and the information we can establish from them. Moreover this concerns the practical evaluation in terms of accuracy and Brier score.

Finally: in the exam you may be asked to give the meaning of a certain concept, but that doesn't mean you have to be able to literally recite definitions or theorems. Concerning the proofs in the syllabus and course slides: you do not need to be able to repeat these; those treated in class were just to refresh and illustrate the use of rules from probability theory, or to show what would break down if underlying assumptions are validated.

October 26, 2018: On Monday, the evaluation form for this course will become available in Caracal: please use it to provide us with feedback about the course. The deadline is November 18.

October 17, 2018: In class some students noted that Blackboard was showing a grade 0.0 for assignment C. I have checked the grades that I entered for those students that notified me and the registered grades are what they should be. I therefore have no idea why Blackboard isn't showing them correctly. If you suspect Blackboard is showing you the wrong grade, please check with me (I cannot see what you're seeing, so I cannot check this myself)!

The grades for assignment C are available in Blackboard's Gradecenter. I will bring the marked assignments to class today (and Friday), so you can have a look at your own work.

Two errors that were made often:

• C3: A Bayesian network graph is by definition an I-map and from an I-map you can never conclude dependency!
• C14/15: many of you did not report the correct independence assumption underlying a naive Bayes network in C14. If the graph in C15 fitted the assumption in C15, then you did receive points for C15, even if it wasn't a naive Bayes network. Since the network will again be used in Assignment D, please verify that you are (now) doing the right thing!

October 9, 2018: The grades for assignment B are available in Blackboard's gradecenter. I will bring the marked assignments to class tomorrow, so you can have a look at your own work.

3 students had incorrect numbers for B5 (correct numbers (rounded): 0.3898 and 0.24645) and therefore should check their Car diagnosis network and the inference algorithm they use; their student numbers are:

5986591, 6602630 and 3930394

Some students did not fill out these numbers at all, so they are advised to verify their model/algorithm is correct as well!

At this stage, everyone should be able to easily anser questions B1 through B5; if you didn't manage to obtain a passing grade (6 or higher) for the assignment ask yourself whether this is due to a lack of understanding or time! B6, asking for a proof, is difficult for most of you, and therefore was considered a bonus question.

September 27, 2018: For those who had indicated that they'd prefer a paper copy of the syllabus and haven't gotten one yet: new copies are expected to be available through A-Eskwadraat by Monday.

August 2018: Most information for the new acadmic year is now available. For information that is still missing (e.g. course slides), see last year's pages.

Learning goals (2018-2019): upon completing this course, the student

1. recognises and understands the strengths and weaknesses of probabilistic graphical models in general and Bayesian networks in particular;
2. understands the relation between probabilistic independence and the graphical representations thereof, and is able to draw conclusions from this relation;
3. understands and is able to apply probabilistic inference in Bayesian networks;
4. has knowledge and understanding of methods for constructing the Bayesian network graph for actual applications;
5. has knowledge and understanding of various methods for quantifying Bayesian networks, including parameterized models (noisy-or), together with their benefits and limitations;
6. understands and is able to apply techniques for evaluating the robustness and quality of a Bayesian network.