Double Growing Neural Gas for Disease Diagnosis

Abstract

We present here a variant of B. Fritzke’s GNG (Growing Neural Gas) — Double GNG (DGNG). In each insertion step of the GNG only one new cell is inserted in the middle of the edge connecting Maximum Resource Vertex (MRV) and a MRV in its direct topological neighbourhood. But in our DGNG two new cells are inserted at the same time. Our goal is to speed up the convergence of the learning process. Although multiple growing cell mechanism can reduce the required number of learning epochs, it can also lead to an increased network structure. Simulation results on some neural network benchmarks indicate that, for many data sets, when the number of new cells in each insertion step is within three, the total performance of networks is at best, but that increasing the number of new cells beyond three has very little benefit and sometimes degrades performance. In this paper we consider only the DGNG. With two disease diagnosis benchmarks (Wisconsin breast cancer and soybean disease diagnosis) we tested the DGNG and indicated that DGNG performs better than the original GNG, measured by the required number of epochs and CPU times.