Distributed Regression: an Efficient Framework for Modeling Sensor Network Data

Carlos Guestrin, Peter Bodik, Romain Thibaux, Mark Paskin, and Samuel Madden

    
         
What you see
   This is a video of the temperature measured over a period of several days (around October 28th) by 48 wireless sensors at the Intel-Research lab in Berkeley, CA. The dots are the actual temperature measurements, and the surface is the regression given by the sensors (this is simulation, but with real data).    
         
Why it is interesting
 

Even though represented using a small number of parameters, the surface approximates reasonably well the temperature at each point. Indeed, it is arguably a better estimate of the temperature than the raw sensor measurements because it cleans out the noise. You can sometimes see a point or two swing widely: these are sensors becoming faulty for an instant.

The point of our IPSN paper (here) is to give a robust and efficient distributed algorithm to compute this surface. Not only is it easier for the user to retrieve a picture of the current temperature, but the sensors all share this global information about their environment and can make deductions such as fault detection.

   
         
The Video
   [Download]