A method for the design of evolutionary multilayer neural networks

A method is presented for the design of evolutionary multilayer neural networks. This method supports both in-depth and in-width evolution. The proposed multilayer structure is obtained by cascading several sections of two-layer nets. The construction of each two-layer section is formulated as a basis selection problem. By constraining the basis elements to be selected from the input parts of the given sample I/O pairs, a solution can be obtained by using a greedy basis selection algorithm. Simulation results are included to illustrate the approach