Home | Administrative Policies | Course Structure | Materials | Assignment Schedule | FAQ

The goal of this project is to use Dynamic Bayesian Networks to try and retrieve information from sensor streams.

 

• This is an individual project. Your work should be done on your own.

• Part One
  • Scan the documentation for GMTK
    • GMTK runs on Linux
  • Get GMTK up and running
  • Here is a sample HMM with files for GMTK
    • sample
• Part Two
  • Create a model for the following environment and specification.
    • 
      • Outside is 0
      • Room A is 1
      • Room B is 2
      • Room C is 3
      • Room D is 4
      • Room E is 5
    • There are "three" people that could work in this office
      • "Alice", "Bob", "Unknown"
      • Alice works in office A
      • Bob works in office D
      • Unknown wanders about randomly and uniformly
      • Only one person is in the building at a time.
    • There are 5 time periods during the day
      • In the morning Alice and Bob both go to their office and sometimes go to the printer room "E"
      • In the late morning Alice and Bob print out more stuff than in the early morning
      • At noon both Alice and Bob leave the building for lunch then return
      • In the early afternoon both Alice and Bob attend a meeting in office B
      • At the end of the day Alice and Bob both go home.
    • The building has a motion sensor which reports the room that the building occupant is in.
      • The motion sensor is mostly accurate but sometimes goes off when a person is in an adjacent room
    • The occupants can't go through walls.
  • Use the following structure
    • 
  • Turn in a copy of your conditional probability tables.
  • Use GMTK to figure out the identity of the person in the following data sets (time,sensor)
    • data set A
    • data set B
    • data set C
• Part Three
  • Hold all parameters fixed in your model except the P(Who) parameters.
    • Train your model using the datasets A,B,C.
      • Doing this should cause your model parameters to drift toward believing that only one person is ever in the building. Who is that ?
    • Hand craft some data sets that are representative of the other person.
    • Retrain your model on data sets A,B,C and your hand-crafted data sets.
      • Continue to hand craft data and redo the training process until you get the model to converge on a 50/50 chance of Alice or Bob being in the building.
      • Turn in a qualititative description of what your training data sets were and how much of it you had to make.
    • This command line might help save you some time:
      • gmtkEMTrain -strFile part3.txt -of1 observationMaster.txt -ni1 2 -fmt1 ascii -inputMasterFile inputMasterFile.txt -outputMasterFile outputMasterFile.txt -objsNotToTrain objsNotToTrain.txt -maxEmIters 5