ObtainingNewDatasets

We have a set of algorithms that process data and employ statistical
methods to look for emerging patterns. The algorithms produce a
model, called a Bayesian Network, that is a succinct high-level
representation of a sequence of data. We are convinced of the broader
applicability of this technique to other kinds of data sets. We’d
like to see what happens if we run other kinds of data through this
machinery.

This technique will work on any sequence of time-stamped event data.
For example:
* Low-level motion data such as positions and orientations of the
user in the environment measured over time.
* High-level motion data: “user enters room A”, ”...enters hallway
B”, ”...leaves floor C”, ”...passes through doorway D”
* The record of a user’s interactions in an environment: “picked
up object A”, “pushed button B”, “opened menu C”

Basically it works on a collection of data-points consisting of “what event happened” and “what time it happened”. The events may low-level, high-level, or anything in between. The events may be reduntant, simultaneous

In our own research we’ve used it on Quake3 data to observe low-level
player motions (“move forward”, “strafe left”, “rotate clockwise”),
and environment interactions (“grab item A”, “jump”, “get shot”).
We’ve also used it to view patterns in prose text (considering each
character to be an event, and considering the character’s position in
the text to be the timestamp).

There are several possible outcomes of these analyses:

• A representation of all the ways in which users may behave within a
  particular environment. If the data is taken over a sufficiently
  large group of users, it produces a model of the possibility space of
  observed behavior in the environment.

• A statistical classifier. Each trial, or sequence of data (for
  instance, player 300’s performance on map M during a single
  capture-the-flag game”), maps to a particular bayesian model. Two
  baesian models can be compared numerically, and the proximty score can
  be tested for significance w.r.t. the distribution of the bayesian
  models resulting from a particular group of trials. For instance, all
  players who have ever played map M, or the set of trials involving
  players who are ranked in the Top 10. In the context of analyising
  text, we might compare a particular sentence of unknown origin with
  the set of all sentences written by Tolstoy.