Monday, January 12, 2009

Weighted averaging from Sasha's paper

Each data epoch (1.024s), 16 epochs form a sweep.
estimate of the variance of the epoch
weighting factor w(j) = (1/var(j))/(sum(1/var(k))
weighted accumulated epoch a(i) = sum(w(j)*x(i))
if the 4th sweep, the number of epochs are 4.
the accumulated sweep with weighted averaging has 16 weighted accumulated epochs.
Each weighted accumulated epoch.

if the 4th sweep, we have to calculate 64 (4 sweeps * 16 epochs/sweep) epochs variances.
The goal is to obtain 16 weighted accumulated epochs.
Basically, weight of each epoch should be 1/var;
The reason to divide it to sum(1/var) is normalization.