•Initial Data Training: no prior knowledge ofdataset
•Using random number as the initial values
•High accuracy requirements
•Using the no-linear least square fitting back-propagationalgorithm
•Levenburg-Marquardt (LM) Back-propagation:
•Accurate, less efficient, not always converge, depending onthe initial conditions.
•Use a modified gradient decent back-propagation togenerate the initial parameters for LM algorithm.
•May need a few runs to get good data training results
•Data Retraining
•Captures the daily changes to the data patterns.
•Small, but significant sometimes.
•Previous states are used as the starting point.
•Requires the efficient algorithms for daily operations
•Using the modified gradient decent back-propagation