Course talk:CPSC522/Elicitation of Factored Utilities

Don't quote the whole motivation; use your own words. Only use a quote when it helps your argument, or you want to talk about someone else's work (e.g., to critique it). It should never introduce the topic. [Note that if you are using someone else's words, you need to quote it, otherwise it is academic misconduct. So adding quotes means that you will not be committing academic misconduct, but it doesn't make for good writing.]

In problem definition how does D relate to X? What does "includes a distribution" mean. I think we need more of a big picture.

Use "For example,..." when introducing an example into the text (as opposed to general statement.)

In the additive utility, I didn't understand "are independent of each other and are comparable with the same unit of reference."

The definition of GAI should be made clearer. Don't give the same definition as for additive utility. Doesn't additive utility also have the property that you ascribe to GAI? Can't additive utility also be put in this canonical form? In what sense is this generalized?

Given your definition of GAI, surely you never need to compare (1 stop, business class) with (2 stops, first class). Why do you need global queries?

By "uncertainty about over user preferences", who is uncertain (the system or the user)? I think we are missing the big picture again. This impacts "the best decision with respect to the prior..." Is this the user's best decision or your guess as to the best decision?

In your running example, which is the one with minimax regret? Should we always prune dominated outcomes?

We need a conclusion.

Minor: only capitalize names in the middle of a sentence (e.g., "Bayesian Elicitation" -> "Bayesian elicitation" because "Bayes" is a name).

Thanks very much for your comments. I will correct them soon. Best regards, Ali

The explanations to each of the pieces of your comment are as follows (I have tried to have the order of my explanations match the order in your comment):

1. I fixed the problem of all the motivation section being quoted from a source. The reason why I used the exact wordings of Braziunas & Boutilier (2008) was that this was the first time I was learning about of Elicitation of Utilities and I thought their explanation of why this is important was very accurate and quite motivating.

2. I have added some more explanation of how $d\in D$ relates to $X$. The relation is that the choice of $d$ will result in a distribution over $X$. I have added an example for how the distribution on $X$ is known when the choice $d$ is made.

3. I added "For example, ..." in all the sections where I introduce an example.

4. The way I understood the difference between the additive and general additive utilities was that the additive utility was a simpler ( and older - in the matter of the time of discovery and analysis) utility function which did not consider the fact that the amount of importance of each attribute is different for different users. Additive utility was actually not really necessary in the page because all consecutive sections assumed the utility function of the user can be represented as a general additive function, so I removed the additive utility section.

5. Again your comment is talking about the difference between additive and general additive utilities, which I myself was not quite sure how we can theoretically differentiate between them so I removed the additive utility section.

6. Your comment says, "Given your definition of GAI, surely you never need to compare (1 stop, business class) with (2 stops, first class)." I think we actually do need to compare (1 stop, business class) with (2 stops, first class) because some people might be willing to trade-off the number of stops with the flight class, but for example, I think that I prefer to have a journey with fewer number of flights even if the flight class is really bad.

Also, I added an example to show why we need global queries.

7. The uncertainty is on the system's side. The system starts with a high uncertainty about the preferences of a user ( and the parameters of their utility function), and by asking them some queries, it tries to model the user's utility function with higher accuracy. I have added some explanations to this section to make it clear.

"The best decision with respect to prior ..." is talking about the prior distribution that the system has learned about the user's utility function and now is trying to find the best decision based on the learned prior distribution. Indeed, the best decision made based on the learned prior distribution may not be the best decision for the user. The difference between the best decision made based on the learned prior and the best decision for the user will decrease if we ask a new query from user.

8. I added the result of doing minimax regret on the running example. Pruning the dominated utilities is one of the criteria in decision making that is popular in the elicitation with feasible utility sets scheme. I think you might have mistaken it with pruning dominated outcomes. Pruning the dominated outcomes is implicitly done in most of the popular criteria - and all of the criteria mentioned in this page - in both the Bayesian elicitation and elicitation with feasible utility sets schemes. Nonetheless, pruning the dominated utilities is not necessarily a good criterion and is definitely not incorporated in the other criteria.

9. Added a conclusion section to the page.

10. Capitalization of letters in the middle of sentence are corrected.

Thanks very much for your help in improving my academic writing skill.

Best regards,

Ali

Overall, I like this page - I think it's well structured and mostly clearly written. Couple of comments:

First, a very minor thing: I think in the problem definition you are a bit inconsistent with using $d$ versus d (ie math-mode versus not). Not a big deal but just noticeable when reading it.

I found the General Additive Utility definition a bit confusing - it's not completely clear how it is different from Additive Utility. When first reading it, I thought that the distinguishing feature was that, in general additive utility, "the user should be indifferent between any two attributes with the same marginals on each factor". After reading it more closely, I think what you're saying is that this property holds for both types of Additive Utility, and the distinguishing feature is expressing it with scaling constants? I think this could be written a little more clearly.

I liked your example comparing the utilities of different flights, but think it would be nice if you gave an example of what the question Q^1 could be. I'm currently struggling to think of what Q^1 could result in e.g. P(Q^1 = yes | u=u_1 ) being 0.8, and not just always 0 or 1.