The Markovian Streams (MStreams) Project now has website: http://mstreams.cs.washington.edu/

Markovian streams are correlated, uncertain, ordered data streams; technically, they’re the output of probabilistic inference on a temporal graphical model.  The goal of the MStreams project is to develop algorithms that allow for efficient management and querying of Markovian streams using techniques from probabilistic data management.

This work is especially exciting from an RFID systems engineering perspective because we can generate and work with MStreams that represent RFID data – and we are.  In our Cascadia system we process RFID streams with particle filters to produce MStreams.  This process  transforms a raw RFID data stream (complete with missing tag reads and gaps from “dead zones”) into a stream containing a smoothed, probabilistic representation of location (i.e. a probability distribution over the person or object’s location at each timestep).  We’re now looking at adapting the MStreams Project’s database engine, Lahar, for use with our probabilistic RFID streams.  This should facilitate expressive queries over RFID data while improving accuracy through probabilistic data management.  More work in this vein soon…