In the second phase of learning, the value
gathered at the root is rebroadcast to all the nodes and
each node reacts, based on its matching state. It may:
• self-delete,
• increment pty , counters, and/or cover_count,
• generalize or cause the new pair to do so,
• attempt to prevent the creation of a new node.
A new node is always created if the training pair
is an exception to some other existing pair. A new
node is also created if the training pair generalizes or
if there does not seem to be a "natural" way for the
current network to predict its output.