README

Time filter experiments. 

take1/ - 

Modeling pulse error as a function of variation in the pulse interval. An 
analysis of one day's worth of data from a fixed transmitter. Since the data
rate (data_rate.png) is not representative of variations higher than 1.0, the 
results here are inconclusive. See take3/ for plans for continued work in this 
direction. 

  Results:

    SCORE_INTERVAL = 3 minutes
    SCORE_NEIGHBORHOOD = 20 seconds
    SCORE_ERROR = lambda(x) : (-0.6324 / (x + 7.7640)) + 0.1255 seconds
    SCORE_THRESHOLD = 0.20 (lower bound on relative score for pos. est.) 


take2/ - 

Qualitatively evaluate the results of take1/. Does the time filter yield 
performance gains over the band filter alone? 


take3/ [planned] - 

Grab a week's worth of data for a fixed transmitter. (Say, deploymentID=102 
for 19/11/14 - 26/11/14.) If the data rate sufficiently resembles pulses for
various pulse variation quantizations, then segment the points into good and
bad as in take1/. Then we'll partition a testing and training set and do it 
up right. (Look into employing Dwork's diff. privacy technique here.)