log in  |  register  |  feedback?  |  help  |  web accessibility
Logo
Reconstructing preferences and priorities from opaque transactions
Friday, October 10, 2014, 11:00 am-12:00 pm Calendar
  • You are subscribed to this talk through .
  • You are watching this talk through .
  • You are subscribed to this talk. (unsubscribe, watch)
  • You are watching this talk. (unwatch, subscribe)
  • You are not subscribed to this talk. (watch, subscribe)
Abstract

There has been significant work on learning about utility functions of single agents from observing their behavior in isolation.  In this talk, I will discuss the problem of learning about utilities of a collection of agents, when the only information available is some kind of overall outcome of a joint interaction or opaque transaction.

For example, consider an auction of a single item where n agents each draw values from their own personal probability distributions D_i, and the only information that can be observed is the identity of the winner (highest bidder).  From repeated observations, plus the ability to enter the auction (and win or lose) yourself, can you reconstruct the relevant parts of the individual distributions?  Or consider a setting with multiple items where agents have combinatorial preferences, and where a seller is running a mechanism that you do not know.  From observing the results of a sequence of these interactions, can you learn both the preferences of the buyers *and* the mechanism of the seller?  In this talk I will discuss algorithms in the context of both of these problems.  In the process we will see connections to decision-list learning in learning theory and Kaplan-Meier estimators in medical statistics.

This is joint work with Yishay Mansour and Jamie Morgenstern

This talk is organized by Jeff Foster