
TYPEKeyword of Integer
Type Indicate type of neural network to be used in calculations.
standard {0} -- feed-forward
neural network trained according to back-propagation TRAINING
with pre-determined architecture and number of hidden layers;
CASCOR (QuickProp1)
{1} -- Cascade Correlation algorithm, updated from the original
version of S.
Falhman by Vasily V. Kovalishyn. The training of weights is done using
the original
version of QuickProp algorithm.
CASCOR (TRAINING)
{2} -- Original implementation (unpublished, Tetko, 1997) of the Cascade
Correlation algorithm with training of weights selected according to
TRAINING option. The structure of Cascade Correlation is completely different
from standard back-propagation algorithm. This algorithm begins training
with no hidden neurons that are added in process of training. Thus, the
structure of Cascade Correlation nets grows during with time of network
training.
The parameter NEURONS for Cascade Correlation
network should contain only one value that indicate maximum allowed
number of hidden neurons. Usually, NEURONS=100
hidden neurons is a good choice. The training of this algorithm is terminated
if RMSE error for the validation set does not decrease after 20 new hidden
neurons were added to network or LIMIT criterion
was satisfied. The default value is {0}.
See FAQ if you have questions. How to cite this applet? Are you looking for a new job in chemoinformatics? |