Collection: "dumb" and/or clever questions for preparation

Disclaimer: Dieser Thread wurde aus dem alten Forum importiert. Daher werden eventuell nicht alle Formatierungen richtig angezeigt. Der ursprüngliche Thread beginnt im zweiten Post dieses Threads.

Collection: “dumb” and/or clever questions for preparation
So I’ve started learning stuff.

First things first: Will any of the topics be NOT relevant for the exam? E.g. I feel lie a lot of “small” topics like dynamic baysian networks or the new logic based learning results haven’t really been discussed in depth and there’s been no exercices or examples on what you could do on paper :'D So can I expect it to not happen in the exam?

  1. Lately in the lecture Mr. Kohlhase kept on talking about “cheating in proofs” I would’ve looked it up, but the slides are’t uploaded yet and I’m confused to what he actually means/does.

  2. Slide 756 says the size of a HMM with 20 variables is 2^20*2^20 - whyyyy? I don’t see that size happening when considering hyperdimensional matricesor stuff.

  3. Slide 830 we have the neat formula as below - can you check the spelling/brackets/whatever on this? Unless - we are not supposed to know :stuck_out_tongue: But currently this looks a bit confuseing with the 2 indexed x and the x squared and shouldn’t it be N-1 sum xi if there’s no difference?

Attachment:
2018-07-06 22_19_19-notes.pdf - Adobe Acrobat Reader DC.png: https://fsi.cs.fau.de/unb-attachments/post_156983/2018-07-06 22_19_19-notes.pdf - Adobe Acrobat Reader DC.png

1 „Gefällt mir“

So am I, but also I have little context as to in which topic he might have said that. Regarding the exam, it’s safe to say that it’s not relevant, but if you’re curious some more context could help me to figure out what he might have meant by that :wink:

Here’s the more exhaustive description of that situation in Russell/Norvig

The point being that a DBN only has arrows between those state variables that are actually causally dependend (which is why it has only three parents even though there are 20 variables per time slice), whereas the HMM has an entry in its matrix for every conditional probability, even those that are potentially zero.

Oh yes, that definitely needs brackets:


@1 Well Mr. Kohlhase repeatedly mentioned it in (Prolog-style? haven’t been arroud for KI1) proofs/proof trees. Where he “cheats” at some point of the branch xD I just didn’t get it in the lecture. It’s supposed to provide logical reasoning of some kind, but how xD?


So, in the context of first-order logic in KI1? I’m sorry, I have no idea what that might refer to, but I’ll make him aware of this thread and see if he can answer :smiley:


Let me try to explain what I think this means (I attended KI1).

Prolog works like this:
You write facts and rules.
You ask Prolog to answer a query.
Prolog tries to resolve the query with the rules and facts you wrote down earlier.
If Prolog can not resolve any further and it has not found a fact, it aborts / throws an exception.
I believe that „Prolog with Cheating“ can somehow introduce facts when they are needed and stores these „cheated facts“ as assumptions, in order to get something else than just an exception.

Example (simple Prolog, so relax :wink: ):

The facts and rules:

human(plato).
human(X):-not lizard(X).
mortal(X):-human(X).

Line 1: plato is human (fact)
Line 2: humans must be not lizards (rule)
Line 3: mortals must be humans (rule)

Now we query for [m]mortal(X).[/m]
So we ask „Who is mortal?“ And get [m]X = plato[/m]

If we now delete Line 1, and query again for [m]mortal(X).[/m] Prolog aborts, since there is no facts about human(X), or in other words we can not resolve X to a constant.

(Prolog Syntax: constants are lower-case, Variables are upper-case)

I believe that „Prolog with Cheating“ would now assume that there is some X, so that human(X) holds and then it would try to find out what condition this X has to fulfil in the first case.

So if we delete Line 1, and query for [m]mortal(X).[/m], I would expect something like [m]for some X where not lizard(X) [/m] from „Prolog with Cheating“.

1 „Gefällt mir“

One more; we have this thing called value functions. Is that the same as utility functions essentially?
The slide stated something akin to “a total ordering on states is a value function (ordinal utility function)” - So it’s just a utility with no two states sharing the same utility?


Yes :slight_smile:


When we are asked to explain how to compute, or to explain the algorithm that computes something, is it enough to show the formulas/equations that would be generated for that computation?


ONLY formulas are not enough, if there’s no explanation what the symbols occurring mean, or how to iterate/update/proceed