Contact information

Lecturers:

for all your questions and remarks related to the contents and examination of this course



Fellow students:

You can discuss the course, its contents and the exercises with your fellow students through the discussion forum on Blackboard.



FAQ

How can I register for the course?

Registration for this course proceeds through OSIRIS. Also note that registration generally closes about 8 weeks before the start of the course (see registration dates). If you are planning on taking this course in your first year, you will be notified how to register during the Master Introduction in the first week. A lecturer has no OSIRIS-access and cannot register you for the course! In case of problems, please contact the student desk of the graduate school of natural sciences.

Can I do a (thesis) project on a topic related to this course?

Sure! See this projects page to see whom in the department does research on related topics and just contact one of them..

Can I get solutions to the exercises?

In contrast to bachelor courses, master courses are not scheduled with exercise classes or practicals. The syllabus contains various exercises for each chapter; for each class, the schedule indicates what useful exercises are for the topics of that lecture. For most types of exercise either the syllabus or the slides include a completely worked out example (note that, especially in the beginning, some of the proofs and examples are in fact exercises from the syllabus). Other exercises can be solved in different ways; a correct outcome usually means that you used a correct approach to solve it. For most exercises the syllabus therefore provides either the outcome or a hint. If you need more help than that, you can ask the lecturer; alternatively, you can ask your fellows students through the discussion forum on Blackboard.

If we post more complete solutions (either in the syllabus, or online) then most of you who make the exercises now will be tempted to only look at the solutions. You could argue that that is your own responsibility, but our experience is that most students need the practice. A lot of exercises, especially those involving Pearl's algorithm, require practice: it seems so simple if you look at completely worked out examples.

Why don't you let us build an actual application?

Well, you're free to do so.... However, the reason for not letting you do this as part of the course is that building a realistic Bayesian network application is not possible within the time frame of the course (this can in fact take years), unless you learn it from a dataset. All problems associated with the latter are not unique to BNs and do not teach you a lot about BNs. If you want to learn models from data we suggest you also take the Datamining course (INFOMDM).

Why don't we get a coding assignment?

Coding algorithms: since we am not interested in testing your programming skills, we are not in favor of such practicals for this particular course. How ever, if you think it is insightful to do so, there's no-one stopping you...

Specifically concerning Pearl's algorithm: it is used to teach you not just how inference works, but more importantly why inference works. For this reason, Pearl is suitable; however, as explained in class, it is hardly used in practice. That is another reason why we don't want to ask you to code Pearl.

Why does a course on probabilistic reasoning focus only on Bayesian networks?

As explained in class, Bayesian networks are basically the only tool for probabilistic reasoning. That is, if we stick to models that correctly define a joint probability distribution, and consider probabilities that adhere to the Kolmogorov axioms. Dropping the latter assumption means deviating from 'standard' probability theory, which is not what we want to do in this course. Dropping the former assumption means looking at systems like rule-based systems, Dempster-Shafer theory, Fuzzy sets, and Probabilistic logics; all such systems to some extent deal with numbers, sometimes even referred to as probabilities, but typically do not operate according to the laws of probability theory within an overall joint distribution. You could argue that, mathematically, these are not as nice and clean as Bayesian networks, but others will disagree. Note that these alternative systems are briefly discussed in the first lecture. Concerning all 'real' statistical models other than Bayesian networks: you can in fact consider lots of them as special cases of Bayesian networks (e.g. Naive Bayes classifiers, Hidden Markov models) or of a continuous version of a Bayesian network (e.g. Kalman filters, Icing models); the latter are not discussed in class though. So although the topic of Bayesian networks seems a limited account of probabilistic reasoning, we pretty much cover a large area. If you're not convinced: google 'probabilistic reasoning'.