Function modeling in sigmoidal and quasi sigmoidal back-propagation

摘要

The features of the back-propagation algorithm using either sigmoidal or quasi sigmoidal sigmoidal activation functions for function approximation problems are compared. Quasi sigmoids are here proposed as a generalization of the standard sigmoidal function which can be shown to be particularly useful to cope with with a selected class of problems. The comparison is carried out in terms of various performance indexes showing that the proposed functions are superior to the standard fixed sigmoids. The reason of that is related to the flexibility of quasi sigmoids which implies a better matching to different kind of problems. The comparison between the two functions also highlights that the use of quasi sigmoids allows for a better choice of the projection directions of the input data. The performance of the networks on some benchmark tasks are analyzed and commented.