A NOVEL ALGORITHM FOR GATED EXPERT NEURAL NETWORKS

Abstract

This thesis presents a new algorithm for modular networks or sometimes called gated expert networks. The network consists of several expert networks and a gating network that decides how the outputs of the expert networks should be combined to form the final output ofthe system.

A new function is introduced for the gating network in which an optimisation theory framework is adopted. Also each expert network has associated with it a center that represents the center of its domain of influence in the input space. Accordingly each expert network learns certain patterns from the input space and becomes an expert in its region. The impact of this gating function on the splitting of the input space is shown. Also, the update of the weights of the expert networks, and the adjustment ofthe centers of the expert networks is illustrated.

A comparison is made between the standard backpropagation and the new algorithm for the gated expert networks. Several simulations were performed on different applications, four of these applications are theoretical problems and the last one is a real-world problem (the Nile River flood). The results showed that there is an improvement that varies between 17% and 64.5%. Also, a computational complexity comparison was done between the standard backpropagation and the new algorithm for the gated expert networks.