Recursive Cascade Correlation and Recursive Multilayer Perceptron, a comparison

M. Hagenbuchner and A.C. Tsoi

An extension of the classical Cascade Correlation Model as proposed by \cite{micheli1,sperduti96} allows the processing of more general data structures. This method is known as recursive cascade correlation (RCC). In \cite{perso1} a recursive model based on MLP is described (RMLP). Both models build and train the network in a very different way. In addition, the network architectures are also different so that a direct comparison of the two models has not been possible. This paper suggests a number of modifications of the RCC and RMLP network architectures so as to obtain a neural network architecture which is of a similar type. These modifications are particularly interesting since the two models RMLP and RCC build up their respective architectures in a very different manner, and hence, it is interesting to learn how these two models compare if the underlying architecture is identical. In addition, experimental results, which are conducted on a real world learning problem, are to show what effect the different weight arrangements in the architecture have on these models.