A practical gated expert network

Difficulties in training multilayer networks for strongly nonlinear problems has led some researchers to propose the gated expert networks. The idea is based on having several local “expert networks”, where each learns a particular region of the input space. A “gating network” combines the outputs of the expert networks to produce the final output. We propose a practical gated expert algorithm and adopt an optimization theory framework. The algorithm slices up the input space and approximates each region using an expert network, in an analogous fashion like the way the spline curve fitting technique approximates each of its regions separately. Simulation results indicate the effectiveness of the proposed approach