LEAST SQUARES, CALIBRATION (JOINT WORK WITH SHAM KAKADE)

Abstract: The problem of on-line prediction is try to predict an arbitrary sequence of numbers about as well as can be done if you had seen all the data ahead of time. Of course, to make this a "fair" problem, we will have to restrict the possible after-the-fact forecasts. This talk will use the easiest possible loss function--namely least squares. I will focus on some new spin-offs of this theory. In particular, it is easy to show that a weakly calibrated forecast exists. Using these strategies, both correlated equilibria and Nash equilibria can be learned in a fictitious play like setting.

Location: 
Elath Hall, 2nd floor, Feldman Building, Edmond J. Safra Campus
Dates: 
Sunday, March 25, 2007 - 16:15 to 18:15
Old Lecturers: 
DEAN FOSTER
Old Lecturers University: 
UNIVERSITY OF PENNSYLVANIA