Computational Intelligence

228 Chapter Six—Neural Network Implementations

on the type of activation function of the output PEs. For a back-propagation

network, the activation function is usually a linear function or one of several

S-shaped functions. The mean-square error is also accumulated for this training

pattern.

The current state transitions to the state

BP_BACK_PROPAGATION_HIDDENS.

Following is the state handler routine.

static void bp_back_propagation_output (void)

{

1int idx;

double tempA,tempB;

for (idx = 0; idx < (bp.arch.layers[bp.arch.size - 1].size); idx++)

{

tempA = (target_out[idx] - bp.arch.layers[bp.arch.size - 1].

neurons[idx].out);

switch (bp.arch.layers[bp.arch.size ...