LMSEstimator update rule implementation wrongly updates the weights taking the LMS cost value into account, this means that if the error is huge (and the cost quadratically grows with it) the weight could get an update value almost proportional to the LMS value (smoothed by learning rate alpha).
The correct update rule should be used which only uses the residual.
The side effect of this bug is that:
- cost may vary a lot
- weights could not converge to a good solution