Rotation - This option selects training patterns in the order they appear in the .PAT file or the .TRN file if Calibration is being used. Use when like training patterns are dispersed evenly throughout the training set.
Random - This option randomly chooses the training patterns, although it does not guarantee that every pattern will be chosen an equal number of times. Use when the training set contains patterns that are cyclical (such as sales figures that follow seasonal variations) and you want the network to give answers independently of clustered data patterns.
Generally speaking, binary and toy problems should use rotational pattern selection and complicated problems with numeric outputs, such as financial problems, should use random selection. These are guidelines rather than rules and a user should try different combinations to see which works the best with a particular application.
Recurrent networks must be trained with rotation, since order is important.