Computational Intelligence

Postprocessing 195

likely to have an effect. Another approach sometimes used is the shuffling of the data

after each epoch (as opposed to random selection). The results of different training

runs, each with randomly or sequentially selected patterns, should be compared for

the effect of presentation order on the outcome of training.

Another important consideration in preparing data for training a neural network

is the addition of noise to perturb the data. By adding noise (jitter) to the data, the

result is a convolutional smoothing of the target (Reed, Marks, and Oh 1995). This is

a technique that may be helpful when only a relatively small number