|
Abstract:
The goal of quite a bit of AI research is to design systems that can make (or at least recommend) decisions on behalf of users, in domains ranging from low stakes problems--like book, music, or movie recommendations--to high stakes decisions--like medical, logistic, or financial advice. To make such decisions, a user's preferences (or utility function) for outcomes must be known. Preference elicitation refers to the problem of finding out (enough) about a user's preferences to make an (approximately) optimal decision on her behalf.
I plan to discuss (or ideally, just lead discussion of) the problem of preference elicitation. For one view of the problem, I suggest looking at the following paper:
Chajewska, U., L. Getoor, J. Norman and Y. Shahar (1998).We'll briefly discuss this paper and it's view of elicitation. (Other interesting papers on the topic can be found here as well.)
Utility Elicitation as a Classification Problem.
In Uncertainty in Artificial Intelligence 14 (UAI '98), pp. 79-88.
Winner of the UAI '98 best student paper award.
I'll then present a model of preference elicitation as a sequential decision problem, specifically, as a partially observable Markov decision process (POMDP), and why I think this is exactly the right way to think about elicitation. I'll then describe some of the bottlenecks--both conceptual and computational--associated with solving such a POMDP, and suggest some ways one might tackle these problems.
There will be no theorems, no empirical results, just a lot of speculation and discussion of a number of (what I think are :-) very interesting research problems. Bring your thinking caps!