|
Onno Zoeter Associate Researcher
Applied Games Group onnoz [@] microsoft.com
|
|
|
Onno Zoeter Associate Researcher
Applied Games Group onnoz [@] microsoft.com
|
|
Research
I am a member of the Applied Games Group (APG) together with Ralf Herbrich,
Thore Graepel, Phil Trelford, and Joaquin Quiñonero Candela. APG's
official drive is to
leverage the methods of approximate probabilistic inference for
addressing relevant applications both in recreational games and in
abstract decision games played in the real world.
Currently I am paricularly interested in the games that are played on the web. I view information extraction tasks on the web such as search, ranking, recommendation, etc. as collaborative filtering games between different types of players. One could look at the PageRank algorithm for instance as a step in this direction where webpage creators are granted "voting rights" in such a filtering game. My research focusses on making these games fair, relevant, and reliable. The current emphasis is on (Bayesian) machine learning techniques for advert-relevance prediction. Research hobbies In my spare time (which is a thing of the past really) I like to think about Bayesian psycho-physics models: in essence coming up with a theoretical framework that describes how "optimal machines" would perform in psycho-physics experiments if they had to rely on the same set of fallible sensors as human subjects.Together with Tom Heskes, Tjeerd Dijkstra, Jan van Gisbergen, and Ronald Kaptein I've tested our notion of "optimal machines" on data from studies of the vestibular system (perception of the direction of gravity in rotated human subjects). We find that the predicted behaviour of the "optimal machines" and the experimental data agree remarkably well. |