FeatureSet method: offset

Syntax: <feature set> offset <feature> <source> [optional flags]

This methods provides an adaptive offset removal (and normalization) for SVector as well as FMatrix features. The mean (mean vector for FMatrix features) is reestimated for each new value x(i):

      mean(i) = alpha * x(i) + (1-alpha) * mean(i-1)
      y(i)    = x(i) - mean(i) 

Same thing can be done for the mean of squares that can be used to calculate the standard deviation.

      smean(i) = alpha * x(i)^2 + (1-alpha) * smean(i-1)
      dev(i)   = sqrt(smean(i) - mean(i)^2)
      y(i)     = ( x(i) - mean(i) ) / ( a * dev(i) ) 

By using the flags -upMean and -upSMean you can update the resulting mean vectors of an FMatrix feature in FVector objects for later use. In runon mode you should give the same FVector object as start value and update name as in the example.

-alpha <float>

Adaptation factor (see example1).

-mean <float_or_FVector>

Start value for mean. If the source feature is FMatrix you can specify an FVector here. This can be a vector definition like {0.1 0.2 ..} or a name of a FVector object. The definition can also be read from a file. Then <FVector> must be: @filename.

-smean <float_or_FVector>

Similar to -mean. A start value can be defined for the mean of squares.

-a <float>

see example 2

-upMean <FVector object>

This must be an FVector object because it will be updated with the last calculated mean vector.

-upSMean <FVector object>

Similar to upMean. Here the mean of the squared values is updated.

-mindev <float>

If the values of the source are getting nearly constant the deviation is very small and dividing by this value is not what was intended by the normalization. So a minimum deviation helps that the resulting values never become too big.