As mentioned before, the weights are arranged in a single linked list of weight structures. This structure contains a unique weight number and control information. Weights can either be fixed, which means not trainable, or variable, which means that they may be updated, see also WEIGHT-flags. The actual weight is stored as a double precision floating point value.
A unique feature of the environment is that the weight structure also holds a pointer to a list of previous weight values. This makes it possible to store the history of the network to study the evolution of the weights during training. General purpose data space allows the user to add own information to the weight structure.
Figure 3.4: The weight structure.