A new approach in multilayer perceptron-forward propagation (FP)

A new feedforward propagation algorithm (FP-1) in multilayer feedforward network is proposed. In this new approach the errors are transferred forward and the established architecture, the fixed neurons and the trained weights previously in network are not changed. Several new definitions are introduced and described as the FP-1 local mapping space, FP-1 areas, nonlinear distribution chart, half FP-1 local mapping space, etc. With this new algorithm, we will know and understand every step of the approximation, and hence control the approximation to reach arbitrary desired precision. An important principle is presented on how to decompose the given mapping into global and local components, and combine with them to reach final desired results. Although there are more to explore in mathematics, this paper focuses on the basic idea of how to realize and implement the proposed algorithm