Function decomposition network

Abstract

Novel neural network architecture is proposed to solve the nonlinear function decomposition problem. Top-down approach is applied that does not require prior knowledge about the function's properties. Abilities of our method are demonstrated using synthetic test functions and confirmed by a real-world problem solution. Possible directions for further development of the presented approach are discussed. © 2009 Springer Berlin Heidelberg.