August 20, 2008, 14.30 hrs

Sander Evers

Abstract:

In our sensor data lab, the Bluetooth scanners detect devices that have been in the neighborhood at some point during a 10-second scan period.

When modeling this sensor data, it is more accurate to keep this notion of an interval, rather than to pinpoint the data at one specific instant; it is also useful for the integration of sensor data with different time granularities. In this talk, we show how to process such interval-valued data in a probabilistic way, using a variant of the Hidden Markov Model, without having a serious impact on performance.

To ease the analysis of this probabilistic processing, we use a novel technique called a sum-factor diagram. Using this diagram, we can visually rewrite probabilistic queries without worrying at all about the underlying probabilistic semantics. The method also has interesting connections to conventional database query optimization.