Course talk:CPSC522/Value of Information and Control

I think this is a very well written and organized page. The fact that we just learned about value of information in the class made reading it even more smooth for me.

I saw two very minor issues which I think are not mentioned in the previous comments: In the "Value of Control" section, I think using the word "similar to" instead of "like" is better. Also, in the last sentence of the page (the sentence right above "Annotated Bibliography") the part of the sentence after ", and the importance of ..." should be changed to contain a verb.

Another minor point which I did not quite understand is in the "Control with Parents Maintained" section. I actually did not understand what it means to maintain all the parents. In the example from the page on decision networks, Weather is not connected to Umbrella, but when you have converted Forecast to a controlled variable, a link has been added from Weather to Umbrella. I think a little more explanation of what maintaining the parents means will make me clear about this section.

Thanks for pointing out the "like" and the "Control with Parents Maintained" section, which definitely wasn't well explained. I've changed them now.

Here are some comments on the first draft.

I think there needs to be more motivation. Surely the notion of value of information (and control) should be independent of the particular representation. I would like to see a definition of the value of information before you should how it can be computed from a decision network. And also explain why we might be interested.

It would be good to go beyond the example used in the textbook. (The story even in this example seems incomplete; why didn't you include the value of the forecast for the case when you can observe the weather? This would make clear some of the issues.) There are also subtle issues for sequential decisions about when the information is obtained that you could try to explain.

You should number your figures, and the the numbers in the text. (Resizing and printing moves the figures around).

Thanks for the comments, I've made those changes now

First off, this is a very clearly written entry, with illustrative examples.

A minor question: Is there any reason that you use a lower case x for a random variable in the Definition of Value of Information?

Another thing is that the value of imperfect information is defined as a sub-optimal value due to some noise. In the section on the value of perfect information, it sounds like that the value of information of weather forecast is imperfect, compared with that of weather itself. So, can the information being indirect count as noise in this context?

Hi, thanks for the feedback.

I think the lower case x is a mistake because I was combining the notation from different sources - thanks for pointing it out.

Yes, I think 'weather forecast' is essentially a noisy version of 'weather'. So getting perfect information about the 'weather forecast' is like getting imperfect information about 'weather'. I'll look at changing this to make it clearer